arXiv:2603.08655v1 Announce Type: new Abstract: We introduce OfficeQA Pro, a benchmark for evaluating AI agents on grounded, multi-document reasoning over a large and heterogeneous document corpus. The corpus consists of U.S. Treasury Bulletins spanning nearly 100 years, comprising 89,000 pages and over 26 million numerical values. OfficeQA Pro consists of 133 questions that require precise document parsing, retrieval, and analytical reasoning across both unstructured text and tabular data. Frontier LLMs including Claude Opus 4.6, GPT-5.4, and Gemini 3.1 Pro Preview achieve less than 5% accuracy on OfficeQA Pro when relying on parametric knowledge, and less than 12% with additional access to the web. When provided directly with the document corpus, frontier agents still struggle on over half of questions, scoring 34.1% on average. We find that providing agents with a structured document representation produced by Databricks' ai_parse_document yields a 16.1% average relative performance gain across agents. We conduct additional ablations to study the effects of model selection, table representation, retrieval strategy, and test-time scaling on performance. Despite these improvements, significant headroom remains before agents can be considered reliable at enterprise-grade grounded reasoning.

arXiv:2603.08706v1 Announce Type: new Abstract: Training large language models (LLMs) as autonomous agents often begins with imitation learning, but it only teaches agents what to do without understanding why: agents never contrast successful actions against suboptimal alternatives and thus lack awareness of action quality. Recent approaches attempt to address this by introducing self-reflection supervision derived from contrasts between expert and alternative actions. However, the training paradigm fundamentally remains imitation learning: the model imitates pre-constructed reflection text rather than learning to reason autonomously. We propose Agentic Critical Training (ACT), a reinforcement learning paradigm that trains agents to identify the better action among alternatives. By rewarding whether the model's judgment is correct, ACT drives the model to autonomously develop reasoning about action quality, producing genuine self-reflection rather than imitating it. Across three challenging agent benchmarks, ACT consistently improves agent performance when combined with different post-training methods. It achieves an average improvement of 5.07 points over imitation learning and 4.62 points over reinforcement learning. Compared to approaches that inject reflection capability through knowledge distillation, ACT also demonstrates clear advantages, yielding an average improvement of 2.42 points. Moreover, ACT enables strong out-of-distribution generalization on agentic benchmarks and improves performance on general reasoning benchmarks without any reasoning-specific training data, highlighting the value of our method. These results suggest that ACT is a promising path toward developing more reflective and capable LLM agents.

arXiv:2603.06590v1 Announce Type: cross Abstract: The Abstraction and Reasoning Corpus (ARC) is designed to assess generalization beyond pattern matching, requiring models to infer symbolic rules from very few examples. In this work, we present a transformer-based system that advances ARC performance by combining neural inference with structure-aware priors and online task adaptation. Our approach is built on four key ideas. First, we reformulate ARC reasoning as a sequence modeling problem using a compact task encoding with only 125 tokens, enabling efficient long-context processing with a modified LongT5 architecture. Second, we introduce a principled augmentation framework based on group symmetries, grid traversals, and automata perturbations, enforcing invariance to representation changes. Third, we apply test-time training (TTT) with lightweight LoRA adaptation, allowing the model to specialize to each unseen task by learning its transformation logic from demonstrations. Fourth, we design a symmetry-aware decoding and scoring pipeline that aggregates likelihoods across augmented task views, effectively performing ``multi-perspective reasoning'' over candidate solutions. We demonstrate that these components work synergistically: augmentations expand hypothesis space, TTT sharpens local reasoning, and symmetry-based scoring improves solution consistency. Our final system achieves a significant improvement over transformer baselines and surpasses prior neural ARC solvers, closing the gap toward human-level generalization.

arXiv:2603.06594v1 Announce Type: cross Abstract: Automated \enquote{LLM-as-a-Judge} frameworks have become the de facto standard for scalable evaluation across natural language processing. For instance, in safety evaluation, these judges are relied upon to evaluate harmfulness in order to benchmark the robustness of safety against adversarial attacks. However, we show that existing validation protocols fail to account for substantial distribution shifts inherent to red-teaming: diverse victim models exhibit distinct generation styles, attacks distort output patterns, and semantic ambiguity varies significantly across jailbreak scenarios. Through a comprehensive audit using 6642 human-verified labels, we reveal that the unpredictable interaction of these shifts often causes judge performance to degrade to near random chance. This stands in stark contrast to the high human agreement reported in prior work. Crucially, we find that many attacks inflate their success rates by exploiting judge insufficiencies rather than eliciting genuinely harmful content. To enable more reliable evaluation, we propose ReliableBench, a benchmark of behaviors that remain more consistently judgeable, and JudgeStressTest, a dataset designed to expose judge failures. Data available at: https://github.com/SchwinnL/LLMJudgeReliability.

arXiv:2603.06620v1 Announce Type: cross Abstract: The growing demand for automated graph algorithm reasoning has attracted increasing attention in the large language model (LLM) community. Recent LLM-based graph reasoning methods typically decouple task descriptions from graph data, generate executable code augmented by retrieval from technical documentation, and refine the code through debugging. However, we identify two key limitations in existing approaches: (i) they treat technical documentation as flat text collections and ignore its hierarchical structure, leading to noisy retrieval that degrades code generation quality; and (ii) their debugging mechanisms focus primarily on runtime errors, yet ignore more critical logical errors. To address them, we propose {\method}, an \textit{agentic hierarchical retrieval-augmented coding framework} that exploits the document hierarchy through top-down traversal and early pruning, together with a \textit{self-debugging coding agent} that iteratively refines code using automatically generated small-scale test cases. To enable comprehensive evaluation of complex graph reasoning, we introduce a new dataset, {\dataset}, covering small-scale, large-scale, and composite graph reasoning tasks. Extensive experiments demonstrate that our method achieves higher task accuracy and lower inference cost compared to baselines\footnote{The code is available at \href{https://github.com/FairyFali/GraphSkill}{\textcolor{blue}{https://github.com/FairyFali/GraphSkill}}.}.

arXiv:2603.06642v1 Announce Type: cross Abstract: Test-Time Training (TTT) language models achieve theoretically infinite context windows with an O(1) memory footprint by replacing the standard exact-attention KV-cache with hidden state ``fast weights'' W_fast updated via self-supervised learning during inference. However, pure TTT architectures suffer catastrophic failures on exact-recall tasks (e.g., Needle-in-a-Haystack). Because the fast weights aggressively compress the context into an information bottleneck, highly surprising or unique tokens are rapidly overwritten and forgotten by subsequent token gradient updates. We introduce SR-TTT (Surprisal-Aware Residual Test-Time Training), which resolves this recall failure by augmenting the TTT backbone with a loss-gated sparse memory mechanism. By dynamically routing only incompressible, highly surprising tokens to a traditional exact-attention Residual Cache, SR-TTT preserves O(1) memory for low-entropy background context while utilizing exact attention exclusively for critical needles. Our complete implementation, training scripts, and pre-trained weights are open-source and available at: https://github.com/swamynathanvp/Surprisal-Aware-Residual-Test-Time-Training.

arXiv:2603.06816v1 Announce Type: cross Abstract: The alignment problem refers to concerns regarding powerful intelligences, ensuring compatibility with human preferences and values as capabilities increase. Current large language models (LLMs) show misaligned behaviors, such as strategic deception, manipulation, and reward-seeking, that can arise despite safety training. Gaining a mechanistic understanding of these failures requires empirical approaches that can isolate behavioral patterns in controlled settings. We propose that biological misalignment precedes artificial misalignment, and leverage the Dark Triad of personality (narcissism, psychopathy, and Machiavellianism) as a psychologically grounded framework for constructing model organisms of misalignment. In Study 1, we establish comprehensive behavioral profiles of Dark Triad traits in a human population (N = 318), identifying affective dissonance as a central empathic deficit connecting the traits, as well as trait-specific patterns in moral reasoning and deceptive behavior. In Study 2, we demonstrate that dark personas can be reliably induced in frontier LLMs through minimal fine-tuning on validated psychometric instruments. Narrow training datasets as small as 36 psychometric items resulted in significant shifts across behavioral measures that closely mirrored human antisocial profiles. Critically, models generalized beyond training items, demonstrating out-of-context reasoning rather than memorization. These findings reveal latent persona structures within LLMs that can be readily activated through narrow interventions, positioning the Dark Triad as a validated framework for inducing, detecting, and understanding misalignment across both biological and artificial intelligence.

arXiv:2603.06862v1 Announce Type: cross Abstract: Artifact Evaluation (AE) is essential for ensuring the transparency and reliability of research, closing the gap between exploratory work and real-world deployment is particularly important in cybersecurity, particularly in IoT and CPSs, where large-scale, heterogeneous, and privacy-sensitive data meet safety-critical actuation. Yet, manual reproducibility checks are time-consuming and do not scale with growing submission volumes. In this work, we demonstrate that Large Language Models (LLMs) can provide powerful support for AE tasks: (i) text-based reproducibility rating, (ii) autonomous sandboxed execution environment preparation, and (iii) assessment of methodological pitfalls. Our reproducibility-assessment toolkit yields an accuracy of over 72% and autonomously sets up execution environments for 28% of runnable cybersecurity artifacts. Our automated pitfall assessment detects seven prevalent pitfalls with high accuracy ($F_1$ > 92%). Hence, the toolkit significantly reduces reviewer effort and, when integrated into established AE processes, could incentivize authors to submit higher-quality and more reproducible artifacts. IoT, CPS, and cybersecurity conferences and workshops may integrate the toolkit into their peer-review processes to support reviewers' decisions on awarding artifact badges, improving the overall sustainability of the process.

arXiv:2603.06974v1 Announce Type: cross Abstract: We present Elenchus, a dialogue system for knowledge base construction grounded in inferentialist semantics, where knowledge engineering is re-conceived as explicitation rather than extraction from expert testimony or textual content. A human expert develops a bilateral position (commitments and denials) about a topic through prover-skeptic dialogue with a large language model (LLM) opponent. The LLM proposes tensions (claims that parts of the position are jointly incoherent) which the expert resolves by retraction, refinement, or contestation. The LLM thus serves as a defeasible derivability oracle whose unreliability is structurally contained by the expert's authority. Our main technical contribution is a mapping from Elenchus dialectical states to material bases in Hlobil and Brandom's NonMonotonic MultiSuccedent (NMMS) logic, satisfying Containment and enabling the elaboration of logical vocabulary that makes explicit the inferential relationships negotiated in the dialectic. We demonstrate the approach on the W3C PROV-O provenance ontology, where a single dialogue session elicits and structures design tensions that a domain expert can articulate, corresponding to decisions documented in a retrospective analysis of the ontology's design. Using pyNMMS, an automated NMMS reasoner, we verify that the structural properties of the resulting material base (nontransitivity, nonmonotonicity, and independence) correspond to specific PROV design rationales, demonstrating end-to-end integration from dialogue through formal reasoning.

arXiv:2603.06976v1 Announce Type: cross Abstract: We present the first large-scale, cross-domain evaluation of document chunking strategies for dense retrieval, addressing a critical but underexplored aspect of retrieval-augmented systems. In our study, 36 segmentation methods spanning fixed-size, semantic, structure-aware, hierarchical, adaptive, and LLM-assisted approaches are benchmarked across six diverse knowledge domains using five different embedding models. Retrieval performance is assessed using graded relevance scores from a state-of-the-art LLM evaluator, with Normalised DCG@5 as the primary metric (complemented by Hit@5 and MRR). Our experiments show that content-aware chunking significantly improves retrieval effectiveness over naive fixed-length splitting. The top-performing strategy, Paragraph Group Chunking, achieved the highest overall accuracy (mean nDCG@5~0.459) and substantially better top-rank hit rates (Precision@1~24%, Hit@5~59%). In contrast, simple fixed-size character chunking as baselines performed poorly (nDCG@5

arXiv:2603.07017v1 Announce Type: cross Abstract: Safety alignment is critical for deploying large language models (LLMs) in real-world applications, yet most existing approaches rely on large human-annotated datasets and static red-teaming benchmarks that are costly, difficult to scale, and slow to adapt to evolving model behaviors. Moreover, overly conservative safety mechanisms can reduce model usefulness by rejecting sensitive but legitimate queries. We introduce Self-MOA (Self Multi-Objective Alignment), a fully automated framework for aligning small language models using weak supervision from automated evaluator models. Self-MOA operates as a closed loop that dynamically generates model-specific red team prompts, constructs preference data from model-generated responses, and aligns models via multi-objective preference optimization to jointly optimize for safety and helpfulness. Across multiple small language models and safety benchmarks, Self-MOA achieves a 12.41\% improvement in safety while preserving helpfulness, using as little as 11 times less training data than human-supervised alignment baselines. These results demonstrate that adaptive, automated alignment can reduce the dependence on static, human-curated safety pipelines in resource-constrained settings.

arXiv:2603.07023v1 Announce Type: cross Abstract: Despite the promise of Retrieval-Augmented Generation in grounding Multimodal Large Language Models with external knowledge, the transition to extensive contexts often leads to significant attention dilution and reasoning hallucinations. The surge in information density causes critical evidence to be submerged by voluminous noise, which complicates the discernment of relevant fragments within a dense input. In this paper, we propose \textbf{Hit-RAG}, a multi-stage preference alignment framework designed to resolve these cognitive bottlenecks through a progressive optimization pipeline. Our approach systematically refines the utilization of external evidence via three distinct stages. First, Supervised Fine-tuning establishes baseline context awareness to minimize information neglect. Next, Discriminative Preference Alignment enhances robustness against misleading distractors. Finally, Group-Relative Policy Optimization stabilizes logical synthesis to prevent reasoning collapse. Extensive evaluations on eight benchmarks demonstrate that Hit-RAG consistently yields substantial performance gains, enabling models to bridge the gap between context acquisition and accurate reasoning while surpassing much larger counterparts in long-context scenarios.

arXiv:2603.07084v1 Announce Type: cross Abstract: Reward hacking is a form of misalignment in which models overoptimize proxy rewards without genuinely solving the underlying task. Precisely measuring reward hacking occurrence remains challenging because true task rewards are often expensive or impossible to compute. We introduce Countdown-Code, a minimal environment where models can both solve a mathematical reasoning task and manipulate the test harness. This dual-access design creates a clean separation between proxy rewards (test pass/fail) and true rewards (mathematical correctness), enabling accurate measurement of reward-hacking rates. Using this environment, we study reward hacking in open-weight LLMs and find that such behaviors can be unintentionally learned during supervised fine-tuning (SFT) when even a small fraction of reward-hacking trajectories leak into training data. As little as 1\% contamination in distillation SFT data is sufficient for models to internalize reward hacking which resurfaces during subsequent reinforcement learning (RL). We further show that RL amplifies misalignment and drives its generalization beyond the original domain. We open-source our environment and code to facilitate future research on reward hacking in LLMs. Our results reveal a previously underexplored pathway through which reward hacking can emerge and persist in LLMs, underscoring the need for more rigorous validation of synthetic SFT data. Code is available at https://github.com/zohaib-khan5040/Countdown-Code.

arXiv:2603.07111v1 Announce Type: cross Abstract: The Werewolf Game is a communication game where players' reasoning and discussion skills are essential. In this study, we present a Werewolf AI agent developed for the AIWolfDial 2024 shared task, co-hosted with the 17th INLG. In recent years, large language models like ChatGPT have garnered attention for their exceptional response generation and reasoning capabilities. We thus develop the LLM-based agents for the Werewolf Game. This study aims to enhance the consistency of the agent's utterances by utilizing dialogue summaries generated by LLMs and manually designed personas and utterance examples. By analyzing self-match game logs, we demonstrate that the agent's utterances are contextually consistent and that the character, including tone, is maintained throughout the game.

arXiv:2603.07138v1 Announce Type: cross Abstract: Emotion Recognition in Conversation (ERC) is critical for enabling natural human-machine interactions. However, existing methods predominantly employ categorical or dimensional emotion annotations, which often fail to adequately represent complex, subtle, or culturally specific emotional nuances. To overcome this limitation, we propose a novel task named Emotion Transcription in Conversation (ETC). This task focuses on generating natural language descriptions that accurately reflect speakers' emotional states within conversational contexts. To address the ETC, we constructed a Japanese dataset comprising text-based dialogues annotated with participants' self-reported emotional states, described in natural language. The dataset also includes emotion category labels for each transcription, enabling quantitative analysis and its application to ERC. We benchmarked baseline models, finding that while fine-tuning on our dataset enhances model performance, current models still struggle to infer implicit emotional states. The ETC task will encourage further research into more expressive emotion understanding in dialogue. The dataset is publicly available at https://github.com/UEC-InabaLab/ETCDataset.

arXiv:2603.07146v1 Announce Type: cross Abstract: Enabling question answering over tables and databases in natural language has become a key capability in the democratization of insights from tabular data sources. These systems first require retrieval of data that is relevant to a given natural language query, for which several methods have been introduced. In this work we present and study a table retrieval mechanism devising fine-grained typed query decomposition and global connectivity-awareness (DCTR), to handle the challenges induced by open-domain question answering over relational databases in complex usage contexts. We evaluate the effectiveness of the two mechanisms through the lens of retrieval complexity which we measure along the axes of query- and data complexity. Our analyses over industry-aligned benchmarks illustrate the robustness of DCTR for highly composite queries and densely connected databases.

arXiv:2603.07368v1 Announce Type: cross Abstract: Biases in large language models (LLMs) often manifest as systematic distortions in associations between demographic attributes and professional or social roles, reinforcing harmful stereotypes across gender, ethnicity, and geography. This position paper advocates for addressing demographic and gender biases in LLMs through a dual-pronged methodology, integrating category-theoretic transformations and retrieval-augmented generation (RAG). Category theory provides a rigorous, structure-preserving mathematical framework that maps biased semantic domains to unbiased canonical forms via functors, ensuring bias elimination while preserving semantic integrity. Complementing this, RAG dynamically injects diverse, up-to-date external knowledge during inference, directly countering ingrained biases within model parameters. By combining structural debiasing through functor-based mappings and contextual grounding via RAG, we outline a comprehensive framework capable of delivering equitable and fair model outputs. Our synthesis of the current literature validates the efficacy of each approach individually, while addressing potential critiques demonstrates the robustness of this integrated strategy. Ensuring fairness in LLMs, therefore, demands both the mathematical rigor of category-theoretic transformations and the adaptability of retrieval augmentation.

arXiv:2603.07372v1 Announce Type: cross Abstract: Quality Estimation (QE) is essential for assessing machine translation quality in reference-less settings, particularly for domain-specific and low-resource language scenarios. In this paper, we investigate sentence-level QE for English to Indic machine translation across four domains (Healthcare, Legal, Tourism, and General) and five language pairs. We systematically compare zero-shot, few-shot, and guideline-anchored prompting across selected closed-weight and open-weight LLMs. Findings indicate that while closed-weight models achieve strong performance via prompting alone, prompt-only approaches remain fragile for open-weight models, especially in high-risk domains. To address this, we adopt ALOPE, a framework for LLM-based QE that uses Low-Rank Adaptation with regression heads attached to selected intermediate Transformer layers. We also extend ALOPE with recently proposed Low-Rank Multiplicative Adaptation (LoRMA). Our results show that intermediate-layer adaptation consistently improves QE performance, with gains in semantically complex domains, indicating a path toward more robust QE in practical scenarios. We release code and domain-specific QE datasets publicly to support further research.

arXiv:2603.07394v1 Announce Type: cross Abstract: Visual Question Answering (VQA) is a core task for evaluating the capabilities of Vision-Language Models (VLMs). Existing VQA benchmarks primarily feature clear and unambiguous image-question pairs, whereas real-world scenarios often involve varying degrees of ambiguity that require nuanced reasoning and context-appropriate response strategies. Although recent studies have begun to address ambiguity in VQA, they lack (1) a systematic categorization of ambiguity levels and (2) datasets and models that support strategy-aware responses. In this paper, we introduce Ambiguous Visual Question Answering (AQuA), a fine-grained dataset that classifies ambiguous VQA instances into four levels according to the nature and degree of ambiguity, along with the optimal response strategy for each case. Our evaluation of diverse open-source and proprietary VLMs shows that most models fail to adapt their strategy to the ambiguity type, frequently producing overconfident answers rather than seeking clarification or acknowledging uncertainty. To address this challenge, we fine-tune VLMs on AQuA, enabling them to adaptively choose among multiple response strategies, such as directly answering, inferring intent from contextual cues, listing plausible alternatives, or requesting clarification. VLMs trained on AQuA achieve strategic response generation for ambiguous VQA, demonstrating the ability to recognize ambiguity, manage uncertainty, and respond with context-appropriate strategies, while outperforming both open-source and closed-source baselines.

arXiv:2603.07449v1 Announce Type: cross Abstract: Enterprises commonly deploy heterogeneous database systems, each of which owns a distinct SQL dialect with different syntax rules, built-in functions, and execution constraints. However, most existing NL2SQL methods assume a single dialect (e.g., SQLite) and struggle to produce queries that are both semantically correct and executable on target engines. Prompt-based approaches tightly couple intent reasoning with dialect syntax, rule-based translators often degrade native operators into generic constructs, and multi-dialect fine-tuning suffers from cross-dialect interference. In this paper, we present Dial, a knowledge-grounded framework for dialect-specific NL2SQL. Dial introduces: (1) a Dialect-Aware Logical Query Planning module that converts natural language into a dialect-aware logical query plan via operator-level intent decomposition and divergence-aware specification; (2) HINT-KB, a hierarchical intent-aware knowledge base that organizes dialect knowledge into (i) a canonical syntax reference, (ii) a declarative function repository, and (iii) a procedural constraint repository; and (3) an execution-driven debugging and semantic verification loop that separates syntactic recovery from logic auditing to prevent semantic drift. We construct DS-NL2SQL, a benchmark covering six major database systems with 2,218 dialect-specific test cases. Experimental results show that Dial consistently improves translation accuracy by 10.25% and dialect feature coverage by 15.77% over state-of-the-art baselines. The code is at https://github.com/weAIDB/Dial.