arXiv2606.24959asi15 hr ago Stefan Haas, Luca Killmaier, Alireza Javanmardi, Eyke Hüllermeier
Abstract: Ordinal classification (OC) arises in high-stakes domains such as medicine and finance, where uncertainty quantification must account for the severity of ordinal errors. Conformal prediction (CP) provides distribution-free prediction sets with marginal coverage guarantees; however, its practical effectiveness depends critically on the choice of nonconformity function. We introduce a CP method for ordinal classification based on the ranked probability score (RPS), a proper scoring rule defined over cumulative predictive distributions. Although it reflects ordinal risk quite naturally, it has largely been neglected in conformal ordinal prediction (COP). When used as a measure of nonconformity, RPS yields median-centered contiguous prediction sets by construction. The method is model-agnostic, supports both assessed and grouped ordered categorical outcomes, and permits efficient implementation compared to greedy interval selection procedures. Across multiple ordinal image and tabular datasets, RPS-based CP produces contiguous prediction sets and strikes a favorable balance between prediction set width and the magnitude of ordinal miscoverage relative to existing CP methods.
Jun 26, 20261 views
arXiv2606.24956asi15 hr ago Ranhui Yan, Jia Cai, Mengzhu Chen, Haodong Yang
Abstract: Spectral graph neural networks (GNNs) interpret message passing as frequency-selective filtering. While low-order spectral filters are efficient, their limited selectivity often leads to weak attenuation outside the passband, whereas high-order alternatives introduce optimization challenges. We propose DCQ-GNN, a spectral GNN based on a compact bank of adaptive convex--concave quadratic filters. By restricting the filter order to two while explicitly exploiting complementary curvature, DCQ-GNN improves spectral selectivity as quantified by Dirichlet energy and entropy measures without resorting to high-order polynomial expansions. The model fuses filter outputs through a node-adaptive gating mechanism to enable node-wise structure-aware spectral selection. We provide a formal spectral analysis grounded in Dirichlet energy attenuation, von Neumann entropy, and curvature polarity, and derive explicit characterizations of filter behavior across varying levels of homophily and structural perturbations. Extensive benchmarks on 10 datasets show that DCQ-GNN ties for the top average rank (3.0) on heterophilic graphs and obtains the second-best rank (4.2) on homophilic graphs, remaining competitive with representative high-order polynomial spectral filters. Furthermore, under strong structural perturbations, DCQ-GNN exhibits substantially smaller performance degradation compared to both first-order and high-order baselines. These results demonstrate that curvature-aware quadratic banks provide a robust and efficient alternative to high-order spectral models while preserving optimization stability and computational efficiency.
Jun 26, 20261 views
arXiv2606.24948asi15 hr ago Randhir Kumar
Abstract: Knowledge graph embedding (KGE) models predict single-hop links well but have no mechanism for zero-shot compositional queries: multi-hop questions whose relation chains never appeared during training. Holographic Reduced Representations (HRR), which bind and unbind symbols via circular convolution, are a theoretically attractive candidate, since binding is approximately invertible and associative. We test whether this promise holds. We study two holographic memory variants, real-valued HRR and phase-only Fourier HRR (FHRR), each with a modern Hopfield cleanup, on FB15k-237 over five seeds. Four findings follow. First, both are competitive single-hop retrievers (filtered MRR 0.358 +/- 0.002 for HRR, 0.350 +/- 0.021 for FHRR). Second, neither composes zero-shot: accuracy stays at chance across all cleanup temperatures. Third, the main contribution, we localise the failure mechanistically. A hop-1 probe shows the memory recovers the correct intermediate entity with high fidelity (MRR 0.896 +/- 0.002 for HRR), yet composition still fails even with a verified-correct intermediate. A second probe shows why: posing the ground-truth second-hop fact as a standalone atomic query, bypassing composition entirely, already recovers it at only 0.26 to 0.48x average atomic accuracy, uniformly across relation fan-out. The bottleneck is not the bind-unbind algebra or the cleanup; it is that facts compositional chains pass through are intrinsically harder for the superposed memory to retrieve, a capacity and interference effect present already at a single hop. Fourth, we prove (Lemma 4.1) that FHRR's softmax cleanup is not phase-equivariant, compounding the primary failure on the minority of chains where hop-1 itself errs. Fixing zero-shot composition requires improving retrieval capacity under superposition, not just redesigning the cleanup.
Jun 26, 20261 views
arXiv2606.24937asi15 hr ago Haggai Roitman
Abstract: The Hitchhiker's Guide to Agentic AI is a comprehensive practitioner's reference for building autonomous AI systems. The book covers the full stack from first principles to production deployment, organized around a central thesis: building great agentic systems requires understanding every layer of the pipeline, not just one. The book opens with the LLM substrate -- transformer architecture, GPU systems, training and fine-tuning (SFT,LoRA, MoE), model compression, and inference optimization -- treated as essential foundations rather than the primary focus. It then develops the alignment and reasoning layer: reinforcement learning from human feedback (RLHF), PPO, DPO and its variants, GRPO, reward modeling, and RL for large reasoning models including chain-of-thought and test-time scaling. The second half is devoted to agentic AI proper. Topics include agentic training and trajectory-based RL, retrieval-augmented generation (RAG and Agentic RAG), memory systems (in-context, external, episodic, and semantic), agent harness design and context management, and a taxonomy of agent design patterns. Inter-agent coordination is covered in depth: the Model Context Protocol (MCP), agent skills and tool use, the Agent-to-Agent (A2A) communication protocol, and multi-agent architectures spanning centralized, decentralized, and hierarchical topologies. The book concludes with agent development frameworks, agentic UI design, evaluation methodology for agentic tasks, and production deployment. Each chapter pairs rigorous theoretical foundations with implementation guidance, code examples, and references to the primary literature.
Jun 26, 20262 views
arXiv2606.24902asi15 hr ago Arnesh Banerjee, Ayushi Bhattacharjee
Abstract: The "First Proof" benchmark [1] posed ten research-level mathematics questions to the strongest publicly available LLMs and found them consistently wrong-not silent, but confidently, fluently wrong. This paper asks why. Working from the per-question post-mortems in First Proof's Appendix A, I identify four failure modes: citation fabrication (F1), premise smuggling (F2), silent problem reformulation (F3), and local-to-global compatibility gaps (F4). I then audit eight one-shot proofs generated by Gemini 2.5 Flash on Questions 1, 2, and 5 of the benchmark, using two instruments built specifically to surface F1 and F2. The central finding is uncomfortable for anyone who sees retrieval-augmented generation (RAG) as the obvious fix: not one of the eight proofs contained a confirmed fabricated citation, yet every single one contained at least one load-bearing claim asserted as a "fundamental result" or "standard argument" with no justification attached. That failure mode-F2, premise smuggling-is invisible to citation verification by design. A premise-audit instrument I introduce flags it at 100% precision (5/5 judge-confirmed flags are true positives) and 50% proof-level recall in this corpus. The taxonomy and the audit together suggest that the right long-term objective is building inference-time pipelines that prevent these failure modes from occurring, not just detecting them after the fact. Index Terms--Large language models, mathematical reasoning, hallucination, premise smuggling, failure-mode taxonomy.
Jun 26, 20261 views
arXiv2606.24901asi15 hr ago Hao Jiang, Enneng Yang, Guojie Zhu, Yibin Chen, Yunkun Xu, Zifu Kou, Jiayi Li, Chong Chen, Zhao Cao, Li Shen
Abstract: Continual learning capability is critical for Industrial LLMs, as deployed models must be continuously updated to meet evolving requirements and environments, rather than repeatedly retrained from scratch. However, most existing research focuses on improvements on static benchmarks, failing to capture real industrial needs. In this survey, we reformulate Industrial Continual Learning (ICL) for LLMs as a closed-loop update-and-release problem in a versioned ecosystem, where updates propagate hierarchically to industrial, application-specific models and LLM-powered applications, with capability inheritance and transfer across versions and model families. From this ecosystem perspective, we identify three core challenges: repeated adaptation erodes model plasticity, foundation-model upgrades break capability inheritance, and long-term sustainability is constrained by deployment requirements. We then organize the technical landscape of ICL around five lifecycle design principles: preserving plasticity headroom, treating upgrades as capability transfer, enabling trustworthy continual reinforcement learning, making training recipes self-optimizing, and building accountability as a base layer for long-term iteration. For each principle, we synthesize representative technical directions. Finally, we evaluate the maturity of each principle and its technical components via an evidence-based lens, identify key gaps hindering real-world deployment, and outline a practical ICL deployment blueprint and a pathway for feeding industrial realities back into academic research.
Jun 26, 20263 views
arXiv2606.24900asi15 hr ago Andrea Mattia Garavagno, Edoardo Ragusa, Paolo Gastaldo, Antonio Frisoli, Claudio Loconsole
Abstract: This paper proposes a new approach to near-sensor computing, in which a lightweight Neural Architecture Search (NAS) is performed directly on the deployment device to find the best tiny neural architecture for analyzing the real-time data acquired through sensors. This new adaptation capability can be particularly useful in the case of human-machine interfaces for which the neural network analyzing the biometrical data can be re-designed each time the user changes, after a guided data collection procedure, fighting the typical data variations between individuals on a new level. To implement the proposed approach a new NAS has been designed and then validated on the Italian Sign Language dataset (ISL), a collection of surface electromyography (sEMG) signals of the signs of the Italian alphabet, using several embedded systems. Moreover, further validation on the Case Western Reserve University dataset (CWRU), a benchmark for intelligent fault diagnosis, is presented to suggest another possible application of the proposed approach. When run on a Raspberry Pi 4, the proposed NAS performs beyond the state of the art proposing a tiny neural architecture having 0.63 times less RAM occupancy and 5.96 percentage points of more accuracy in the case of the ISL dataset; and 0.44 times less RAM occupancy and 0.2 percentage points of more accuracy in the case of the CWRU dataset.
Jun 26, 20261 views
arXiv2606.24898asi15 hr ago Rituraj Sharma, Tu Vu
Abstract: Looped language models turn hidden states into runtime state: each state is decoded for prediction and fed back into future computation. This creates a basic supervision question: which state variables does cross-entropy actually control? We show that dense per-loop cross-entropy controls the variables exposed by the readout, not every variable active in the recurrent transition. Hidden-state scale gives a concrete failure mode. Scale-invariant readouts such as RMSNorm and LayerNorm hide radial scale from the immediate cross-entropy loss, while pre-norm residual recurrence continues to carry and update that same scale. Thus per-loop loss can make early exits usable without controlling recurrent scale. In 44M and 129M looped transformers without inter-loop normalization, per-loop cross-entropy through RMSNorm readouts still drives final hidden-state norms into the thousands or tens of thousands. Scale-visible readouts and explicit norm penalties keep norms in the tens, and scale-removing recurrence is the complementary architectural fix. The resulting design rule is simple: dense supervision trains exits; recurrent scale control requires either making scale visible to a loss or removing it from the loop. Consistent with this rule, scale-controlled variants achieve lower perplexity at matched inference-depth operating points in our variable-depth benchmarks.
Jun 26, 20262 views
D. R. Halvorsen M. Okonkwo L. Petrova J. S. Almeida
Abstract: Most practitioners still operate large language model agents the way they operated chatbots: one prompt, one reply, a human in the loop on every turn. We argue this is a transitional posture, not an endpoint. As agents acquire the ability to discover their own work, isolate their changes, and check each other, the unit of engineering shifts from the prompt to the loop: a scheduled, self-driving cycle that the operator designs once and supervises rarely. We present a field playbook for designing such loops. We decompose a loop into five movements, Discovery, Handoff, Verification, Persistence, and Scheduling, and define five levels of operator autonomy from manual prompting to fully autonomous operation. We give a structural account of why an agent that grades its own work tends to inflate its score, and why an independent verifier, the “thing that can say no,” is the single highest-leverage component in the design. We close with a fourteen-step adoption path and a set of failure modes observed in practice. The central claim is simple: you stop prompting the agent, and start building the system that prompts it.
Jun 25, 20262 views
A Field Study of Designing Loops That Run Themselves
Abstract: Abstract—Over the past two years a string of “XX Engineering”
terms has tracked the pace of model releases. This note examines
the newest of them, Loop Engineering, a term independently
surfaced in June 2026 by Peter Steinberger, Boris Cherny, and
Addy Osmani, and named in writing by Osmani. Unlike prompt,
context, or harness engineering, loop engineering does not teach
the practitioner to do the work better; it removes the practitioner
from the position of doing the work at all. We define the term,
place it as a fourth layer above the harness, and decompose a
single turn of a loop into five moves—discovery, handoff, ver-
ification, persistence, and scheduling—and the six parts that
realize them. We give particular attention to the generator/eval-
uator separation: empirically, an agent asked to grade its own
output tends to praise it, and tuning an independent skeptical
evaluator is far more tractable than making a generator criti-
cal of its own work. We survey three loops running in practice,
from one engineer’s morning triage to Stripe’s enterprise-scale
pipeline merging over 1,300 machine-written pull requests per
week, and we catalog four costs that accrue silently—verification
debt, comprehension rot, cognitive surrender, and token blowout.
We close with a concrete recipe for building a first loop. The
central claim is that loops make generation nearly free and leave
judgment as the scarce resource; the same loop, built by two
people, can yield opposite outcomes.
Index Terms—Agentic AI, software engineering, autonomous
agents, coding agents, generator–evaluator, scheduling, automa-
tion.
Jun 25, 20269 views
arXiv2606.24855asi1 day ago Negin Raoof, Richard Zhuang, Marianna Nezhurina, Etash Guha, Atula Tejaswi, Ryan Marten, Charlie F. Ruan, Tyler Griggs, Alexander Glenn Shaw, Hritik Bansal, E. Kelly Buchanan, Artem Gazizov, Reinhard Heckel, Chinmay Hegde, Sankalp Jajee, Daanish Khazi, Emmanouil Koukoumidis, Xiangyi Li, Hange Liu, Shlok Natarajan, Harsh Raj, Nicholas Roberts, Ethan Shen, Nishad Singhi, Michael Siu, Ashima Suvarna, Hanwen Xing, Patrick Yubeaton, Robert Zhang, Leon Liangyu Chen, Xiaokun Chen, Steven Dillmann, Saadia Gabriel, Xunyi Jiang, Anurag Kashyap, Boxuan Li, Yein Park, Minh Pham, Sujay Sanghavi, Lin Shi, Ke Sun, Yixin Wang, Zhiwei Xu, Erica Zhang, Siyan Zhao, Wanjia Zhao, Jenia Jitsev, Alex Dimakis, Benjamin Feuer, Ludwig Schmidt
Abstract: Agentic language models dramatically expand the applications of AI yet little is publicly known about how to curate training data for broadly capable agents. Existing open efforts such as SWE-Smith, SERA, and Nemotron-Terminal typically target a single benchmark, leaving open the question of how to train models that generalize across diverse agentic tasks. The OpenThoughts-Agent (OT-Agent) project addresses this gap with a fully open data curation pipeline for training agentic models. We conduct more than 100 controlled ablation experiments to systematically investigate each stage of the pipeline, yielding insights on the importance of task sources and diversity. We then assemble a training set of 100K examples from our pipeline and fine-tune Qwen3-32B on this dataset, which yields an average accuracy of 44.8% across seven agentic benchmarks and a 3.9 percentage point improvement over the strongest existing open data agentic model (Nemotron-Terminal-32B, 40.9%). Moreover, our training data exhibits strong scaling properties, outperforming alternative open datasets at every training set size in compute-controlled comparisons. We publicly release our training sets, data pipeline, experimental data, and models at this http URL to support future open research on agentic model training.
Jun 25, 20264 views
arXiv2606.24841asi1 day ago Ahmad Pouramini, Hesham Faili
Abstract: Prompt-based learning has emerged as a dominant paradigm in natural language processing. This study explores the impact of diverse pre-training objectives on the performance of encoder-decoder pre-trained language models across generation and question answering tasks, with a focus on commonsense knowledge retrieval and completion. We highlight the benefits of incorporating multiple objectives during both pre-training and fine-tuning stages. We introduce the Match Task to Objective (MTO) framework and methods for determining the appropriate objective for a given task. This framework offers automated methods to prepare task-related data for adaptation through unsupervised training, based on the identified objective. In the fine-tuning stage, we design novel templates that align with the objectives of the pre-training and adaptation stages. When aligned with task requirements, these strategies can achieve a performance gain of over 120\% compared to conventional methods in few-shot settings. They significantly outperform related works in few-shot settings and exceed the baseline even in full-dataset scenarios. Furthermore, we extend this approach to include prompt-tuning methodologies, providing guidance for more effective soft prompt engineering and optimization. Our strategies significantly enhance prompt-tuning performance as well. These insights hold substantial value, precisely guiding the selection and optimization of models customized for specific tasks. Code is available at this https URL
Jun 25, 20263 views
arXiv2606.24839asi1 day ago Tian Zheng, Kai-Tai Hsu
Abstract: Agentic data analysis systems produce rich outputs, including code, numerical results, and verbal diagnostics. This makes them more challenging to evaluate than single-turn LLM responses. It is therefore necessary to distinguish genuine disagreement between an agent's output and a ground-truth answer from grading artifacts. We investigate how reliably automated graders assess such a system and what strategies improve grading quality by applying LAMBDA, a multi-agent data-analysis system, on 153 numerical QRData tasks from DSGym. We develop and evaluate a three-layer human-AI grading cascade: strict regex matching, LLM-based lenient grading, and snippet-based human inspection, which combines non-GenAI and GenAI strategies with different failure profiles. Both automated graders achieve 100% observed precision (0/70 false positives). The lenient grader's recall is 97% against human labels. A keyword-anchored extraction pipeline raises the strict grader's recall by 60 percentage points over a last-number heuristic; the lenient grader is architecturally parser-independent. An iterative nudge mechanism raises grading run success from 36% to 97% and lenient-pass rates from 16% to 46%; comparing nudging with and without original-question re-injection shows that re-injection offers no benefit, confirming the nudge as an answer template cue. We further observe in this case study that variable type is the task metadata field most consistently associated with grading pipeline dynamics and observed outcome grades.
Jun 25, 20262 views
arXiv2606.24832asi1 day ago Sander Beckers
Abstract: Over a series of seven papers, Andreas & Günther have introduced seven definitions of actual causation and have classified them as belonging to three different, competing, types of accounts: factual difference-making, counterfactual difference-making, and regularity-based. I show that their most recent - factual difference-making - definition instantiates all three types, thereby proving that these are distinctions without a difference. I further compare their novel account to the other six accounts on several crucial examples, revealing that this undermines all seven of their accounts.
Jun 25, 20261 views
arXiv2606.24808asi1 day ago Zidu Liu, Florian Marquardt
Abstract: Quantum computers could outperform classical machines on important problems, but only if the errors that pervade quantum hardware can be corrected at scale. Quantum low-density parity-check (qLDPC) codes offer a promising route to this goal by combining sparse parity checks with finite encoding rate and growing distance, but their construction remains a challenging discrete design problem. Here we introduce structured concept evolution (SCE), a search framework that pairs a large language model with a structured algebraic mutation grammar to discover lifted-product code families, a class of CSS qLDPC codes. Instead of asking the LLM to design codes from first principles, SCE evolves structured concepts consisting of algebraic specifications paired with executable programs that realize them, using hierarchical mutations that modify the group algebra, protograph geometry, or base space. Running SCE, we discover a diverse set of competitive code families, ranging from abelian constructions to families over non-abelian groups beyond those underlying standard designs such as bivariate-bicycle codes, and characterize them under code-capacity depolarizing noise with BP+OSD decoding. These results are obtained with lightweight models (GPT-5.4-mini and GPT-5.4-nano).
Jun 25, 20263 views
arXiv2606.24790asi1 day ago Anand Kamat, Daniel Blake, Brent M. Werness
Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse tasks, yet they remain prone to generating hallucinations. Detecting these hallucinations is critical for deploying LLMs reliably in high-stakes applications. We present Grad Detect, a gradient-based approach for predicting hallucinations by analyzing layer-wise gradient patterns from a single forward-backward pass during inference. Our method shows that the internal gradient structure of a model carries rich information about the correctness of its output. This information is not accessible through output-level signals alone. We evaluate Grad Detect on several Q&A benchmarks across both hallucination detection and model abstention prediction, where it consistently outperforms confidence-based and sampling-based baselines. Through comprehensive layer ablation studies across all eleven models from four architectural families, we find that the final five layers concentrate over 97% of the discriminative gradient signal, enabling efficient deployment with minimal performance loss. Grad Detect provides a unified framework for predicting multiple dimensions of LLM reliability, offering strong predictive performance alongside interpretable insights into where and how model failures originate.
Jun 25, 20261 views
arXiv2606.24752asi1 day ago J. Fernando Hernandez-Garcia, Tomás Figliolia, Beren Millidge
Abstract: The loss of plasticity - the ability of a network to learn new information after having already learned older information - is a fundamental challenge in creating artificial neural networks capable of continual learning. Although this phenomenon has been known for decades, it has mostly been studied in older, relatively small architectures and rarely in natural-language domains. To determine whether loss of plasticity remains a problem in the modern transformer-based LLM paradigm, we study plasticity loss in GPT-style Transformer models trained on a multilingual continual learning problem. Consistent with prior work, we find evidence of plasticity loss across models ranging from 5M to 314M non-embedding parameters, as measured by deterioration on a held-out Vietnamese probing task. We further find that the onset of plasticity loss follows a predictable scaling law, growing sublinearly with model size. These results suggest that larger models may delay the measurable effects of plasticity loss, but that increasing parameter count alone is likely to be insufficient to completely prevent it. We also find evidence of plasticity loss under stationary multilingual training, challenging the view that the phenomenon is exclusive to continual learning with abrupt task changes. Overall, our results suggest that even large Transformer language models trained on natural-language will eventually lose the ability to efficiently adapt to new data after sufficiently long training, in both continual and stationary settings.
Jun 25, 20263 views
Lavinia Ghita, Dhruv Desai, Ioana Boier
Abstract: Large Language Models (LLMs) achieve strong performance across a growing range of domains, yet their scale poses deployment challenges in applications where latency and cost constraints are critical. This paper derives empirical scaling laws for domain-specific LLM compression, quantifying how in-domain and general knowledge performance scale with dataset size, compression ratio, supervision format, and iterative pruning schedule. Using quantitative finance as our application domain, we compare logit-based and LoRA-based distillation under iterative structural pruning, introducing a blended chain-of-thought supervision loss that stabilizes KL-divergence distillation over reasoning traces. In-domain task quality degrades predictably under compression while general-knowledge benchmarks collapse well before the same point; supervision format is the key driver of this tradeoff, with chain-of-thought supervision actively recovering general knowledge that pruning erases. We release the headline dataset FinHeadlineMix, scaling law results, and practical recommendations to provide a reusable framework for domain-specific compression decisions.
Jun 25, 20262 views
arXiv2606.24722asi1 day ago Peter Toth
Abstract: Frontier AI training is increasingly shaped by access to dense, centrally controlled accelerator clusters. This creates a structural advantage for hyperscalers and large centralized laboratories, and makes open or independent AI efforts depend on scarce capital, privileged infrastructure, and data-center geography. We present Spheroid BlockTrain, a decentralized training protocol in which a model is partitioned into independently trainable blocks, each optimized on a local objective derived from the same global target and composed at inference into one model. On byte-level WikiText, BlockTrain reaches cross entropy 1.359 (perplexity 3.89), within about 0.04 CE of a same-setup end-to-end Transformer reference, while each active worker trains only one block and avoids full-model optimizer state. A shared six-worker block training run reaches CE 1.385 by averaging same-block updates into one assembled model. HTTP/TCP transport experiments move real serialized checkpoints and updates, including a public-IP three-host run that improves CE from 5.580 to 1.811 while moving 15.22 GB. For inference, the current BlockTrain path uses one block-stack traversal per full output and serves over direct TCP across three public-network GPU hosts up to a 75.80B-parameter logical fp16 shape, outperforming a matched plain-autoregressive TCP pipeline baseline because it emits a full sequence per WAN pipeline traversal rather than one token per traversal.
Jun 25, 20261 views
arXiv2606.24716asi1 day ago Jonas Klotz, Cassio F. Dantas, Pallavi Jain, Diego Marcos, Begüm Demir
Abstract: Sparse autoencoders (SAEs) are increasingly used to extract interpretable concepts from vision and vision language models, yet existing evaluation methods largely rely on proxy metrics or qualitative inspection rather than measuring semantic correspondence. We present a human-grounded evaluation framework that quantifies alignment between SAE latents and human-annotated concepts, without requiring user studies, and validate this matching through targeted attribute perturbations. To enable this intervention-style evaluation in vision, we construct synCUB and synCOCO, synthetic benchmarks of paired images that differ in exactly one attribute. We introduce Fully-Binary Matching Pursuit (FBMP), a coalition-based matching procedure that supports many-to-one mappings between SAE latents and annotated concepts, and consistently outperforms one-to-one baselines. For functional validation, we propose a Targeted Attribute Perturbation Alignment Score (TAPAScore), which tests whether matched concepts respond selectively and in the expected direction under targeted image-level attribute perturbations. Under sanity checks, our matching and TAPAScore are the only evaluated metrics that reliably distinguish trained SAEs from untrained ones. Across SAEs trained on CLIP and DINOv2 embeddings, we find that increased overcompleteness can reduce perturbation alignment, indicating a reduction in interpretability. Our evaluation framework suggests that moderate dictionary sizes provide the best trade-off, yielding the most interpretable SAEs. Code and datasets are available at this https URL .
Jun 25, 20262 views
arXiv2606.24679asi1 day ago Kunyu Ni, Lei Cao, Jie He, Xiaotong Zhang, Jianfeng Jin, Junyu Dong, Yanwei Yu
Abstract: Data preparation pipelines improve data quality in machine learning by transforming raw tables into learning-ready data through sequential cleaning and feature transformation operators. However, automatically constructing such pipelines is computationally difficult because operator sequences are combinatorial and end-to-end evaluation is expensive. Existing state-of-the-art (SOTA) Multi-DQN methods still face three key limitations: decoupled value estimators weaken long-horizon credit assignment, dataset context is only weakly injected into the policy, and exploration is inefficient in a sparse search space with many invalid states. To address these issues, we propose FlowPipe, a unified framework that formulates pipeline synthesis as conditional probabilistic flow generation over a directed acyclic graph. FlowPipe uses Conditional Generative Flow Networks (C-GFlowNets) with a Trajectory Balance objective to connect terminal validation rewards with early pipeline decisions. It further introduces Deep Semantic Modulation through Feature-wise Linear Modulation (FiLM), allowing LLM-derived logical priors to condition the policy's internal activations according to dataset semantics. In addition, FlowPipe incorporates failure awareness into the flow objective to avoid invalid states and concentrate search on high-potential regions. Experiments on two benchmark suites with 74 real-world datasets show that FlowPipe outperforms SOTA baselines, improving accuracy by 11.96% on average and achieving 12.5x faster training convergence. Source code is available at this https URL .
Jun 25, 20262 views
arXiv2606.24672asi1 day ago Xia Zong, Tuomo Lehtonen, Jussi Rintanen
Abstract: In many decision-making scenarios, acquiring information incurs different costs. We consider the problem of constructing a deterministic evaluation strategy that minimizes the expected cost of evaluating a propositional formula under variable costs and a probability distribution over truth assignments. We present a branch-and-bound algorithm with variable-selection heuristics, pruning, and caching. To the best of our knowledge, it is the first practical exact algorithm for this level of generality. Experiments on random instances demonstrate scalability and quantify the efficiency-quality trade-off of a greedy beam-search variant. We additionally evaluate a structured heart-disease diagnosis instance. Finally, we prove that the problem is $\#P$-hard and contained in $\mathrm{PSPACE}$.
Jun 25, 20261 views
arXiv2606.24669asi1 day ago Kuan-Yen Liu, Ren-Jyun Huang, Ti-Rong Wu
Abstract: Large language models (LLMs) have shown strong potential for planning and sequential decision-making, but prior work often relies on using them as direct controllers, which requires precise action generation and can be unreliable in practice. This paper proposes Latent Action Guidance for Online Reinforcement Learning (LaGO), a framework that uses a pretrained LLM as a latent action prior to softly guide online policy optimization, rather than treating the LLM as an explicit planner or controller. Experiments on both a discrete-control benchmark, CLEVR-Robot, and a continuous-control benchmark, Meta-World, demonstrate that LaGO consistently improves both reward and success rate over Vanilla PPO. In particular, LaGO increases the average success rate from 15.1% to 27.2% on CLEVR-Robot and from 2.7% to 15.2% on Meta-World. Our analysis further shows that stronger pretrained LLMs provide more effective guidance, suggesting that LLM knowledge can improve planning and online decision-making.
Jun 25, 20264 views
arXiv2606.24626asi1 day ago Chenyang Zhu, Jiayu Yao, Kushal Chawla, Youbing Yin, Nathan Wolfe, Pengshan Cai, Jingyu Wu, Spencer Hong, Sangwoo Cho, Shi-Xiong Zhang, Daben Liu, Sambit Sahu, Erin Babinsky
Abstract: As autonomous agents tackle increasingly complex multi-step, multi-agent tasks, their execution trajectories have scaled beyond the constraints of even the largest context windows. Current methods for effectively diagnosing agent failures load the full trajectory into an LLM's context window, which suffers from attention dilution and fails when agentic traces inevitably exceed context limits. To address this, we introduce SAFARI (Scaling long-horizon Agentic Fault AttRibution via active Investigation), a framework that replaces linear context loading with a tool-augmented diagnostic loop. By equipping LLMs with a specialized toolbox to read and search trajectory segments alongside a persistent Short-Term Memory (STM) for cross-turn reasoning, SAFARI effectively decouples diagnostic accuracy from architectural context limits. Our experiments demonstrate that SAFARI outperforms state-of-the-art results by 20% on the Who&When dataset within a 1M token budget, and by 19% on TRAIL GAIA subset on a 25K token budget. Most significantly, SAFARI maintains a 0.58 precision even when the target fault resides 5x beyond the model's native context window, a scenario where traditional evaluators fail entirely.
Jun 25, 20261 views
Andreas Chouliaras, Luke Connolly, Dimitris Chatzpoulos
Abstract: Training safe Reinforcement Learning (RL) systems is inherently challenging, with no guarantee of avoiding unwanted behaviors. The most effective defenses against this are (i) transparency through explainability and (ii) alignment via human feedback. While both show promising results, no publicly available framework currently combines them. To address this, we introduce Themis, an XAI-enabled testing and evaluation framework for Reinforcement Learning from Human Feedback. Themis supports over 200 widely used environments and is easily configurable for experiments in RL, transparency, and alignment. Our results show that Themis can train reward models that match or outperform the environment's true reward signal using human preferences. We also provide a cloud-based platform for collecting human feedback and managing experiments. It is user-friendly, auto-scalable, and supports large participant groups across multiple experiments without extra development overhead. Tests show Themis can support one thousand users in back-to-back experiments on a modest commercial machine.
Jun 25, 20265 views
arXiv2606.24621asi1 day ago Sridhar Mahadevan
Abstract: This paper introduces a categorical account of infinitesimal causality in Frobenius Markov categories equipped with tangent-bundle semantics. IDC captures the infinitesimal layer in which interventions act as tangent deformations of copy/discard structure. Two distinct Frobenius structures interact: (1) the categorical Frobenius algebra on classical variables encoding copying, comparing, and discarding; and (2) the geometric Frobenius integrability condition, namely involutive closure of the intervention distribution, distinct from the algebraic Frobenius structure. Categorical causal sufficiency is defined as the compatibility of these two notions. A key observation is that, for structural causal models, infinitesimal causality is most naturally formulated in the slice of deterministic mechanisms over exogenous variables, with visible stochastic kernels obtained only after pushforward. Interventions are tangent vectors that deform the Frobenius copy/discard operations; their Lie brackets measure whether this deformation preserves classical information-flow structure. Pearl's do-calculus is used as a guiding example of intervention identities: ignoring irrelevant interventions corresponds to counit invariance, action/observation exchange to coproduct compatibility with pushforward, and independence to involutive bracket closure of the visible intervention distribution.
Jun 25, 20261 views
arXiv2606.24618asi1 day ago Michel Leclère, Marie-Laure Mugnier, Guillaume Pérution-Kihli
Abstract: In ontology-based data access (OBDA), multiple data sources are integrated via mappings to an ontology. We consider an OBDA setting based on existential rules and the certain answer semantics. We address the recent issue of query abstraction, which consists of abstracting data queries by translating them to the ontology layer. Since a perfect abstraction may not exist, the notions of minimally complete and maximally sound abstractions have been introduced. We study abstractions within an extension of UCQs with a limited form of inequality and a special predicate marking database constants. While this extension does not lead to an increased complexity of the problems of interest, it is able to express minimally complete abstractions, hence perfect abstractions when they exist. We also characterize maximally sound abstractions by making a new connection with the notion of maximum recovery stemming from data exchange.
Jun 25, 20261 views
Anurag Akula, Satheesh K. Perepu, Abhishek Sarkar, Kaushik Dey
Abstract: Multi-agent reinforcement learning (MARL) addresses the problem of training multiple agents that pursue collaborative, competitive, or mixed objectives. Prior work has investigated transfer learning between source and target domains in MARL; however, the majority of existing approaches impose the constraint that the dimensionalities of the observation space and the global state space must be identical across domains. In this paper, we introduce a method that explicitly accommodates mismatched state-space dimensionalities between source and target domains. The proposed approach, ASALT, incorporates both observation-level and state-level adapters that map the target-domain observations and global states into a shared embedding space, thereby enabling more effective transfer of knowledge across both actors and critics. These adapters can generate embeddings that support efficient strategy transfer across heterogeneous domains. Experimental results on multiple configurations in standard benchmark environments demonstrate that ASALT surpasses existing baselines in terms of sample efficiency and global return in cooperative settings, but its effectiveness depends on the degree of mismatch between source and target domains. Furthermore, our findings indicate that ASALT mitigates negative transfer, which frequently constitutes a major obstacle when transferring policies between domains with differing observation and action spaces.
Jun 25, 20263 views
Yimo Lin, Zhen Zhang, Yibin Li
Abstract: While expert-validated "LLM + script" workflows deliver significant value, they remain static: they encode hard-won domain knowledge yet fail to adapt execution based on feedback. Existing agent research predominantly targets greenfield agents and synthetic benchmarks, leaving the migration of active legacy workflows unresolved. To bridge this gap, we present a reversible, Strangler-Fig migration path that refactors legacy workflows into composable, typed, and auditable stages. Central to this framework is a three-tier convertibility taxonomy (A/B/C), implemented as a routing stage within the system harness, which diagnoses a workflow's readiness and routes it accordingly.
Jun 25, 20268 views
arXiv2606.24551asi1 day ago Xiao Zhou, Siyue Zhang, Yilun Zhao, Jinbiao Wei, Tingyu Song, Arman Cohan, Chen Zhao
Abstract: Computer-use agents can execute software tasks through either graphical interfaces or programmatic command interfaces, but existing evaluations confound interaction modality with differences in tasks, initial states, verifiers, and permitted actions. We introduce a matched execution-layer benchmark of 440 desktop tasks across 18 applications and 12 workflow categories, where screen-only GUI agents and skill-mediated CLI agents receive identical goals, states, and final-state verifiers while being restricted to modality-native actions. In this controlled setting, the strongest GUI agent reaches a 59.1% full pass rate, outperforming the strongest original-skill CLI agent at 48.2%; however, verifier-guided skill augmentation raises CLI success to 69.3%, showing that much of the CLI deficit comes from incomplete skill coverage rather than model capability alone. These results suggest that GUI and CLI expose different execution bottlenecks: GUI agents are limited by reliable grounded interaction over long-horizon workflows, whereas CLI agents are limited by the coverage and scalability of their skill interfaces.
Jun 25, 20261 views
Yanki Margalit, Nurit Cohen-Inger, Erni Avram, Ran Taig, Oded Margalit
Abstract: Multi-agent LLM environments require robust mechanisms for shared knowledge management. This paper formalizes the fleet-memory problem and identifies four foundational failure modes: unauthorized leakage, stale propagation, contradiction persistence, and provenance collapse. To address these, we define explicit systems-level primitives: scoped retrieval, temporal supersession, provenance tracking, and policy-governed memory propagation. These primitives are implemented in MemClaw, a production multi-tenant memory service, and evaluated via ArgusFleet, a reproducible harness testing four governance dimensions. Rather than a baseline comparison, this study measures a live production service, emphasizing real-world architectural insights and negative results. Key Evaluation Results Provenance: Successfully reconstructed 100% of depth-four derivation chains with correct writer identity at sub-second per-hop latency. Propagation: Demonstrated high intra-fleet visibility with zero cross-fleet leakage. Under strong write mode, write-to-visible latency was optimized to a single search round-trip. Production Architectural Issues Discovered Asymmetric Scope Enforcement: Tenant isolation held, but sub-tenant scope was initially bypassed on direct GET-by-id requests for agent-scoped credentials (disclosed and remediated during the study). Pipeline Ordering Conflict: While contradiction supersession works for admitted writes, a synchronous near-duplicate gate can prematurely reject contradictory writes before the asynchronous contradiction detector can evaluate them. Conclusion: Long-context retrieval alone is insufficient for production multi-agent memory. Governed shared memory demands explicit systems-level abstractions, and live evaluation is vital to expose enforcement and pipeline-ordering failures missed by design-only treatments.
Jun 25, 20263 views
arXiv2606.24515asi1 day ago Marta Sumyk, Oleksandr Kosovan
Abstract: Computer-Use Agents (CUAs) execute high-level user goals by perceiving and acting directly within graphical user interfaces. However, reinforcement learning for CUAs remains difficult because open-ended desktop environments rarely provide scalable, machine-readable reward signals: task success is often visually grounded and hard to specify with handcrafted reward functions or dense manual labels. We propose an RL fine-tuning framework that uses autonomous vision-language evaluation as a scalable supervision signal for GUI agents. Given a final screenshot and the original instruction, a Vision-Language Model judges task completion and provides terminal feedback without task-specific heuristics or manual labels during policy optimization. Because autonomous evaluators are imperfect, we model their feedback as a noisy binary reward channel and derive a noise-corrected reward estimator for Proximal Policy Optimization. Experiments across macOSWorld, Windows Agent Arena, and OSWorld show that corrected evaluator rewards outperform both zero-shot baselines and raw evaluator rewards, improving success rates by an average of 12.6 percentage points over zero-shot performance and 5.1 points over raw evaluator fine-tuning. These results suggest that autonomous evaluation can serve as a practical reward signal for RL in GUI environments when evaluator noise is explicitly modeled and corrected.
Jun 25, 20261 views
arXiv2606.24509asi1 day ago Oleg Platonov, Gleb Bazhenov, Dmitry Eremeev, Liudmila Prokhorenkova
Abstract: Due to the wide use of graph-structured data in different fields of industry and science, the development of Graph Foundation Models (GFMs) has recently attracted a lot of attention. While many different types of models are called GFMs, particular interest has been paid to GFMs designed for node property prediction tasks, which is one of the most popular settings in Graph ML with lots of real-world applications from fraud detection in financial and social networks to recommendation systems for e-commerce and user-generated content platforms. While a number of GFMs for this task have been recently proposed, the field has not converged to a unified evaluation setting, and different works evaluate their models in widely different ways, preventing reliable comparison of GFMs with each other and with other types of models. In this work, we conduct a fair and rigorous reevaluation of 9 recent GFMs for node property prediction, comparing them to strong Graph Neural Network (GNN) baselines. We find that, among these GFMs, only the most recent ones based on the Prior-data Fitted Networks paradigm outperform well-tuned GNNs in predictive performance, although at a higher inference cost.
Jun 25, 20261 views
arXiv2606.24506asi1 day ago Zhuoren Ye, Tianyu Wo, Dinghao Xue, Mingming Zhang, Yuchen Teng, Chunming Hu, Renyu Yang
Abstract: Emerging LLM services increasingly host many sparse MoE models, yet most models receive sparse requests and remain cold. This creates a GPU memory problem: model weights are stable and model-determined, while KV-cache is transient and demand-determined. Because cold models rarely reach peak KV-cache demand at the same time, reserving worst-case KV capacity per model wastes memory; a shared KV-cache pool can instead provision aggregate active demand. However, KV-cache sharing is not sufficient when weights and KV-cache remain in a monolithic GPU memory pool. Static weights compete with dynamic KV-cache, and KV-head-limited attention under cold, low-concurrency traffic exposes only a fraction of replicated KV capacity, leading to low GPU memory utilization and weak long-context support. We present CrossPool, a serving engine for cold MoE models that separates FFN weights and KV-cache into two GPU memory pools: a weights pool that consolidates FFN weights across cold models, and a KV-cache pool that dynamically serves active requests while keeping attention local to KV-cache. CrossPool combines a KV-cache planner and virtualizer, a layer-wise pipeline scheduler that hides hidden-state transfers, and persistent kernels with control lowering to reduce CPU-GPU control overhead. With efficient GPU memory pooling, CrossPool underpins bursty long-context requests and outperforms the state-of-the-art kvcached-based multi-LLM serving system, reducing P99 TBT by up to $10.4\times$.
Jun 25, 20261 views
Sauro Succi, Peter V. Coveney, Alex Hansen
Abstract: We discuss reasons why the scaling exponents of current Large Language Models (LLMs) applications are indicating an unsustainable regime in terms of energy resources. We further show that attributing the smallness of such exponents to a numerical bias due to the neglect of a non-zero value of the loss function in the limit of infinite data (``pedestal effect") does not remove the unsustainability issue. Finally, the effects of the smoothness (roughness) of the data on the scaling exponents is commented upon based on an analogy with phenomenological models of fluid turbulence.
Jun 25, 20264 views
arXiv2606.24470asi1 day ago Bojie Li, Noah Shi
Abstract: A real-time agent for general computer use - with games as the most demanding case - must act within tens of milliseconds while still planning over seconds. These two regimes sit at opposite ends of the latency-quality tradeoff. A reasoning VLM (Qwen3-VL-8B-Thinking) deliberates effectively but requires ~1.5 s per response - far too slow for a 15 Hz control loop. In contrast, a reactive VLM (MiniCPM-o 4.5) acts in milliseconds but underperforms on planning-heavy tasks. We couple two frozen models of matched scale (9B reactive, 8B reasoning), leaving the communication channel as the sole trainable component. The standard coupling is a Text Bridge (T): the slow model writes a suffix the fast model reads. We introduce a learned continuous Latent Bridge (L) that projects the slow model's residuals into the fast model's input-embedding space in a LLaVA-style manner, avoiding any text round-trip; both are compared against Fast-Only (F). On 7 Atari games and a driving domain (MetaDrive), tuning the action decoder per channel on held-out seeds, the Latent Bridge matches or beats the Text Bridge in every domain: it significantly improves two games (MsPacman +57%, RoadRunner +28%) and is a safe drop-in elsewhere. Combining both channels interferes destructively (RoadRunner -96%), so only one should be used. The benefit is highly predictable: the bridge helps if and only if slow reasoning already beats fast reaction (T > F) - the Latent and Text gains over Fast-Only move together at r=0.93. MetaDrive is the controlled negative, where the Latent Bridge is demonstrably inert because the Text Bridge adds no value. We release replay recordings and reproducible pipelines.
Jun 25, 20261 views
arXiv2606.24467asi1 day ago Xiaolin Lin, Jingcun Wang, Olga Kondrateva, Yiyu Shi, Bing Li, Grace Li Zhang
Abstract: Long-context large language model (LLM) inference is increasingly constrained by the memory footprint and decoding cost of key-value (KV) caches, limiting sustainable deployment on resource-constrained hardware. Existing KV cache eviction methods typically apply heuristic token scoring over all heads in GQA-based LLMs. These methods ignore the different functionalities of attention heads, leading to the eviction of critical tokens and thus degrading the performance of LLMs. To address this issue, we propose CompressKV, a resource-efficient KV-cache compression framework for GQA-based LLMs. Instead of aggregating attention scores from all heads, CompressKV identifies Semantic Retrieval Heads (SRHs) that capture both the initial and final tokens of a prompt and semantically important mid-context evidence, and uses them to select tokens whose KV pairs should be retained. Furthermore, CompressKV allocates cache budgets across layers according to offline estimates of layer-wise eviction error. Experiments on LongBench and Needle-in-a-Haystack show that CompressKV consistently outperforms existing KV-cache eviction methods across memory budgets. Notably, it preserves over 97\% of full-cache performance using only 3\% of the KV cache on LongBench question-answering tasks and achieves 90\% accuracy with just 0.7\% KV storage on Needle-in-a-Haystack. These results demonstrate an improved resource--performance trade-off for long-context LLM inference. Our code is publicly available at: this https URL
Jun 25, 20261 views
arXiv2606.24453asi1 day ago Theodore Papamarkou, Vladislav Smirnov, Viktor Mazanov, Artem Vazhentsev, Preslav Nakov, Timothy Baldwin, Artem Shelmanov
Abstract: Modern coding agents pair LLM generators with various tools, including cheap diagnostics and expensive verifiers. The tool-use decisions are typically governed by orchestrators that often use fixed rules and ignore uncertainty. We formulate orchestration as cost-sensitive sequential hypothesis testing: a Bayesian controller maintains a belief over candidate correctness and dynamically decides whether to gather more evidence, refine the candidate, verify it, or stop. Across six generators and nine coding benchmarks, Bayesian control proves to be most valuable when verification is costly and critics are informative but imperfect. Beyond control, the belief state yields an interpretable correctness score that outperforms token-probability and raw tool-success baselines for uncertainty quantification.
Jun 25, 20261 views
Heng Ping, Arijit Bhattacharjee, Peiyu Zhang, Shixuan Li, Wei Yang, Ali Jannesari, Nesreen Ahmed, Paul Bogdan
Abstract: Mixture-of-Agents (MoA) architectures improve inference-time scaling by organizing multiple LLM agents into layered reasoning pipelines. However, existing MoA variants fail to sustain gains as depth increases, exhibiting degradation, early plateauing, or saturation. We propose ReM-MoA, a memory-augmented MoA framework that sustains scaling through two mechanisms: (1) a Ranked Reasoning Memory that persistently stores and ranks reasoning traces from all layers using a comparative Reviewer Agent, and (2) a Curated Diversified Memory Routing scheme that exposes different agents to distinct combinations of successful and failed traces, preserving exploration diversity while propagating high-quality reasoning. We further introduce an optional multi-domain Reviewer distillation pipeline that improves ranking quality through frontier-model supervision. Across five reasoning benchmarks spanning math, formal logic, code, knowledge, and commonsense, ReM-MoA consistently outperforms prior MoA variants across both depth and width scaling, and its advantage widens with depth, establishing structured cross-layer reasoning memory as a key missing mechanism for scalable multi-agent inference.
Jun 25, 20262 views
arXiv2606.24414asi1 day ago Andoni Rodriguez, Alberto Pozanco, Daniel Borrajo
Abstract: Formal verification produces machine-checkable certificates that attest to the satisfaction or violation of temporal properties, yet these certificates remain opaque to non-specialist stakeholders. We propose a cycle-consistent neural architecture that generates faithful natural language explanations of verification certificates. A forward network NN1 maps certificates to explanations, and an inverse network NN2 reconstructs certificates from explanations; a symbolic verifier closes the loop, providing a differentiable faithfulness proxy. A pointer-generator mechanism ensures lexical grounding by copying state names directly from the certificate. We evaluate on 420 test certificates spanning six verification methods (bounded proof, k-induction, inductive invariant, lasso, reachability, witness pair) in both YES and NO verdict variants, drawn from a financial compliance domain with 207 named states. Our trained architecture, combined with a hybrid inference-time routing strategy, achieves 90.0% cycle-verified soundness, surpassing a multi- LLM few-shot baseline (76.1% for the best of 16 LLM combinations across four frontier models) by 13.9 percentage points. The neural model wins on 10 of 12 verdict/kind categories, with three categories reaching 100% soundness. The architecture offers 860x faster inference (185 ms vs. 160 s per certificate for the full multi-LLM baseline), offline operation, deterministic outputs, and zero per-inference cost. These results demonstrate that trained specialization outperforms general-purpose LLM prompting for structured certificate explanation, while eliminating the deployment constraints of cloud-based inference.
Jun 25, 20261 views
arXiv2606.24391asi1 day ago Arnaud Ricci
Abstract: We introduce Age of LLM, a turn-based 1v1 benchmark in which two LLMs face off on a 13x7 grid to destroy the enemy base. Three stressors are deliberate: fog of war, full diplomacy (messages, ceasefires, ultimatums; uranium kept secret), and a reliability dimension where every turn must follow a strict JSON schema and an illegal action is silently discarded. The engine is private and each match uses a fresh random map seed and opponent, mitigating the data contamination that affects public benchmarks. Models receive a (near) rule-only prompt with no build-order advice (two tactical seed phrases were present during data collection; see Section 2.7). We benchmark 15 reasoning models across 54 matches and 5,258 actions. Findings: (1) the nuclear rush dominates (78% on the rules-coherent v0.11+ sub-corpus; 85% corpus-wide) with a sole-launcher signature that is largely mechanical under secret-simultaneous launch rules, not a cognitive deterrence failure; (2) military conquest is rare but faster (12.3 vs 18.9 turns); (3) diplomacy is prolific yet almost never consummated; (4) ~58% of illegal actions are fog/state errors, making the illegal-action rate a measure of belief-tracking; (5) -- the least established, and the only one we label exploratory -- a weak link associates reliability with winning. The corpus is small, unbalanced and not side-swapped, so the ranking is a preliminary descriptive view, not a contribution. Beyond ranking, the turn-by-turn traces of actions and messages make the corpus a lens on how LLMs reason under adversarial uncertainty -- their belief-tracking, spontaneous deception, and per-model cognitive "personas" -- which we frame as a future research direction. We release the replay format, an isometric viewer and all replays; engine source on request.
Jun 25, 20261 views
arXiv2606.24381asi1 day ago Shaoshuai Du, Penghao Liang, Yixian Shen, Chuanqi Shi, Hang Zhang, Lun Wang
Abstract: Prompt-based interaction has become a dominant paradigm for using large language models (LLMs), where multiple candidate prompts are evaluated and the top-ranked one is selected for downstream use. This workflow implicitly assumes that prompt rankings are stable under minor variations in evaluation conditions. In this paper, we systematically study prompt ranking stability under common sources of variability, including random seeds and limited evaluation subsets. Across three open-weight LLMs and two benchmark tasks, we find that while overall rank correlations are often moderate to high, the identity of the top-performing prompt frequently changes, leading to unreliable selection decisions. To address this issue, we propose a simple stability-aware selection strategy based on a lower confidence bound, which accounts for both performance and variance. Our results show that this approach improves robustness in unstable settings while remaining competitive in more stable regimes. These findings highlight the importance of accounting for evaluation uncertainty in prompt selection and LLM benchmarking.
Jun 25, 20261 views
arXiv2606.24371asi1 day ago Stefano Mereu, Oleksandr Kuznetsov, Gabriele Marchello, Alessandro Galdelli, Emanuele Frontoni, Adriano Mancini, Ferdinando Cannella
Abstract: Convolutional Kolmogorov--Arnold Networks (KANs) replace the fixed weights of a convolutional kernel with learnable univariate functions. The dominant formulation attaches one such function to every kernel entry and lets it act on pixel values, expressive but parameter-heavy and prone to overfitting. We argue that the learnable functions are better placed in the \emph{structure} of the convolution than on each edge, and we organise the design space along a single axis: whether the function acts on the pixel \emph{values} or on the filter \emph{shape}. We study three realisations. SV-KAN applies one shared univariate function to the values and leaves the spatial filter free and static, aa classical convolution with a single learnable shared activation. AG-KAN keeps the shared value function but supplies the spatial structure through a content-adaptive Gaussian gate. RF-KAN instead moves the learnable functions onto the filter shape, building each filter from oriented ridge profiles expanded in a localised oscillatory (Morlet) wavelet basis with content-adaptive amplitudes. Under a matched four-layer protocol with in-run references and three seeds, RF-KAN and SV-KAN reach $88.47\pm0.10\%$ and $88.20\pm0.31\%$ on CIFAR-10 and $64.40\pm0.19\%$ and $64.57\pm0.30\%$ on CIFAR-100, at about $0.4$M parameters. At this matched scale the shape model and the simplest value model meet at the top, both above a plain convolution and every per-edge KAN we tested, including the official Gram variant, at roughly a fifth of the parameters. A controlled study attributes the RF-KAN gain to an intrinsically localised oscillatory basis and to content adaptivity, and an ablation that removes the learned shape entirely, leaving only the shared value function, collapses accuracy by over forty points, identifying the learned shape as the load-bearing ingredient at this scale.
Jun 25, 20261 views
arXiv2606.24370asi1 day ago Hiroshi Okumura
Abstract: Large language models (LLMs) are increasingly integrated into decision-support roles in business and policy contexts. While prior benchmark studies have primarily evaluated LLMs' causal reasoning capabilities, a more fundamental epistemic dimension has been overlooked: Causal Caution, defined as the propensity to refrain from causal judgment when empirical evidence is insufficient. This study examines the systematic suppression of Causal Caution that occurs when LLMs shift from academic to practical advisory contexts. Using an evaluation rubric inspired by Pearl's Causal Hierarchy (the PCH score), we conducted experiments on four high-performance LLMs -- Claude Sonnet 4.6, Claude Opus 4.7, GPT 5.5, and Gemini 3.1 Pro -- across 480 trials. Causal Caution maintenance rates were 91.7--100.0% in academic contexts but dropped to 6.7--18.3% in practical advisory contexts (Fisher's exact test, p < .001 across all models). Furthermore, when restricted to practical prompts requesting concrete recommendations or explanatory rationales, only 1 of 200 responses (0.5%) maintained Causal Caution. A brief self-correction prompt -- "Please reconsider this judgment from the perspective of causal relationships" -- restored the expression of Causal Caution to maintenance rates of 71.4--100.0% (McNemar's test, p < .001 across all models). These results suggest that helpfulness-oriented response patterns may suppress the expression of Causal Caution in practical advisory contexts, with important implications for organizational governance. The findings indicate that this suppression reflects context-dependent variation in expression rather than an underlying capability limitation, suggesting that multi-agent architectures that separate proposal generation from causal auditing may offer a promising governance design.
Jun 25, 20261 views
arXiv2606.24281asi1 day ago Conor Finlay, Joshua Kurien, Saurabh Dash, Marzieh Fadaee, Beyza Ermis
Abstract: Reasoning language models are increasingly asked not only to answer difficult questions, but also to estimate their likelihood of success. Existing methods typically elicit confidence only once: either before thinking or after answering. We argue that confidence in reasoning models is state-dependent: before thinking, confidence should estimate the chance of the model correctly solving the prompt, while after thinking it should predict whether the realized answer is likely to be correct. This distinction determines the appropriate supervision target: prompt-level success should supervise confidence estimates made after seeing the prompt, while individual answer-level correctness should supervise confidence estimates made after answering. We introduce CALIBER (Calibration Before and After Reasoning), which elicits both estimates and supervises each with the target matched to its information state. Under this unified protocol, CALIBER reduces Expected Calibration Error (ECE) by 52.5% over the strongest single-confidence baseline on BigMathDigits for the 7B model, while achieving the best Brier score and AUROC, and remains within 2.1 points of the best accuracy. Further, on a larger 30B model, CALIBER achieves the best ECE on BigMathDigits while remaining competitive in Brier score and AUROC. Out of distribution, it achieves the best ECE and Brier score on GPQA and TriviaQA, and remains competitive on SimpleQA. Ablations further show that this position-target alignment is most beneficial under distribution shift where it consistently reduces calibration error across all out-of-distribution benchmarks.
Jun 25, 20261 views
arXiv2606.24279asi1 day ago Giovanni Casini, Umberto Straccia
Abstract: In Description Logics (DLs), reasoning under Rational Closure (RC) is a well-known and widely accepted non-monotonic formalism to handle defeasible knowledge. In this paper, we study the application of RC to the core and horn variants of the DL-Lite family of lightweight description logics. We analyze both entitlement (instance checking) and Conjunctive Query (CQ) answering under RC. Our main contribution is providing a plug-in architecture that builds upon existing standard classical reasoners, establishing that reasoning and CQ answering under RC for DL-Lite can be done efficiently with minimal computational overhead.
Jun 25, 20261 views
arXiv2606.24267asi1 day ago Hyunji Nam, Keertana Chidambaram, Dorottya Demszky, Natasha Jaques
Abstract: While in-context learning is generally shown to be effective in Large Language Models (LLMs), bad contexts can cause performance degradation and mode collapse, a phenomenon we call "pigeonholing." **Unintentionally bad** contexts can happen without malicious jailbreaking intents: For example, a user asks the model to justify an incorrect math theorem or fails to correct the model's buggy code. Specifically, we investigate ``pigeonholing" in two scenarios: (1) when the user suggests a solution, and (2) when the conversation context includes the assistant's previous (incorrect) responses. Our experiments across 10 verifiable and open-ended tasks with 10 different models show that pigeonholing manifests in several ways: (1) repeating the incorrect answers from context (leading to 38-40% performance drop), (2) converging on a narrow set of answers in coding and text generation without exploring alternatives, and (3) flipping stance on controversial topics to align with the user or the assistant's previous claims. We find that pigeonholing worsens almost monotonically with the number of conversation turns (performance drops by additional 14+% as repeated mistakes increase from 1 to 5), and pigeonholing-induced mode collapse can happen even when the provided example is correct. As a step toward mitigation, we propose RLVR with synthetic errors which improves models by 43-60% under bad contexts compared to vanilla RLVR baselines.
Jun 25, 20262 views
arXiv2606.24251asi1 day ago Kaiwen Zhou, Constantin Venhoff, Jonathan Michala, Xin Eric Wang, William Saunders
Abstract: Large language models exhibit a growing range of misaligned behaviors such as strategic deception, sandbagging, and self-preservation. As they are increasingly deployed in high-stakes settings, it is critical to reliably detect such behaviors to ensure safe and responsible use. In this work, we propose to monitor misalignment by decomposing it into fine-grained cognitive processes -- misalignment indicators -- and detecting their presence in a model's internal activations via linear probes. We develop a taxonomy of 18 indicators spanning different misaligned behaviors, paired with an automated, meta-plan-guided pipeline that generates multi-turn training conversations. To rigorously evaluate generalization, we construct an out-of-distribution suite combining automated behavioral elicitation, established misalignment benchmarks, and natural benign conversations. Across 5 misaligned behaviors, our probes match a strong LLM judge with 0.935 AUROC on out-of-distribution benchmarks while keeping a low false positive rate on benign traffic. We further perform in-depth analysis to understand the probes and the model's internal representations of misalignment indicators.
Jun 25, 20261 views
arXiv2606.24178asi1 day ago Dominik Lindner, Johann Schmidt, Tom Siegl, Martin Becker, Sebastian Stober
Abstract: Pretrained vision models often misclassify inputs that are rotated, scaled, or sheared, even though these affine transformations leave the object class unchanged. Robustness is usually restored either by building equivariance into the architecture or by retraining with augmentation, both of which require changing or retraining the model. Test-time canonicalization instead leaves the classifier untouched. It undoes the transformation of each input, mapping it to a canonical form near the training distribution before classification. Existing canonicalizers, however, rely on a narrow set of logit-based energy scores and bespoke search procedures, leaving the design space of scoring functions and optimizers unexplored. We reframe canonicalization as out-of-distribution (OOD) detection, which lets any OOD score serve as the energy minimized over transformations. Across benchmarks ranging from handwritten characters and sketches to natural images and 3D point clouds, we systematically evaluate around twenty OOD scores and nine search algorithms, finding that distance-based scores paired with random search and local refinement perform best overall. Because canonicalizing an already-aligned input can hurt accuracy, we add a gated mechanism that transforms an input only when its OOD score indicates this is needed, preserving most in-distribution accuracy while retaining the robustness gains on transformed inputs. Code is available at this http URL .
Jun 25, 20261 views
arXiv2606.24177asi1 day ago Youran Sun, Xingyu Ren, Chugang Yi, Jiaxuan Guo, Kejia Zhang, Jianda Du, Haizhao Yang
Abstract: Large language models are making research production scalable, shifting the bottleneck from producing artifacts to judging claims. We present \textsc{Agon}, a research orchestrator that validates what can be checked inside the workflow and leaves the remaining judgments to human scientists. \textsc{Agon} is built on six design principles: Prompt Economy, Future-Facing, Minimal Prompts, OmniDisciplinary, Massive Parallelism, and Zero-Code. We ran \textsc{Agon} across domains for 444 iterations of Prompt Economy loops, using only small starting topics and no human-written experimental code. These deployments demonstrate scalability while exposing new classes of failure. We organize these failures into a taxonomy along severity, fixability, visibility, and capability locus. The taxonomy separates failures the loops can see and fix from those that require human judgment. Together, these results show that \textsc{Agon} is pushing research toward a new paradigm: machine scales, human steers.
Jun 25, 20262 views