arXiv AI

arXiv:2609.29051v2 Announce Type: replace Abstract: On-policy self-distillation (OPSD) has become a popular recipe for post-training LLM agents. It supervises the agent model at the token level with a stronger teacher view of the same model, obtained by conditioning on privileged information (PI). In this work, we show that in multi-turn agents, this paradigm teaches the student to act with confidence but without the information behind it. The trained agent behaves as if it had privileged information it never observed, and its performance falls well short of plain RL, in the worst case below the untrained base model. Therefore, we propose Privileged Self-Practice (PSP), which keeps the PI and moves it from the loss to the sampler. When the student's rollouts on a task mostly fail, we inject a short per-task instruction written by an analyzer model, sample the task again with the instruction in context, and train on the result with an unchanged GRPO objective. The privileged information stays in the prompt and never enters the loss. Across AppWorld and SWE-bench Verified, with three different student models, PSP obtains the best average score in every setting and is the only method that consistently outperforms plain GRPO, with relative improvements over GRPO of 7% to 58% in average score across settings. Project page: https://sux1ngyu.github.io/PSP/.

arXiv:2609.05227v3 Announce Type: replace Abstract: Recent reports during the AAAI-27 review cycle highlight the risk of reviewers coordinating bids for reciprocal assignment advantage. Prior work treats bidding, reviewer assignment, and review manipulation as separate stages, leaving the lifecycle effects of collusive bidding unclear. Real-world analysis is further constrained by typically unobservable collusive intent and the lack of counterfactuals for the same conference. Motivated by this gap, we introduce \alg, an end-to-end multi-agent simulacra framework for studying reviewer assignment integrity by holding the conference environment fixed and configuring LLM-driven reviewer agents with honest or collusive policies. We further develop an affinity-guided collusive bidding strategy that uses mutual reviewer-paper affinities to construct collusion rings and select target papers, producing expertise-consistent rather than arbitrarily targeted attacks. Controlled experiments show that collusive bidding more than doubles target-paper capture and that assigned colluders score target papers about two points higher than honest co-reviewers, while conference-wide effects remain comparatively modest. Evaluated bid-phase detectors provide only limited evidence of collusion: in a fixed-triplet detector stress test, native positive-bid graphs are confounded by benign affinity, while a Very-High-only diagnostic view enables precise but low-coverage local recovery.

arXiv:2608.27847v2 Announce Type: replace Abstract: Interactive medical diagnosis requires an agent to gather patient evidence through sequential questions before committing to a diagnosis. Large language models (LLMs) now make such consultations feasible, and existing approaches typically select questions by how much they reduce uncertainty among candidate diseases. However, this objective is insensitive to the unequal clinical costs of diagnostic errors and omitted severe alternatives, so minimizing uncertainty need not minimize final diagnostic loss. To bridge this gap, we propose CaRDI (Calibrated Risk for Diagnostic Interaction), a planning framework built around a clinical loss combining primary diagnosis error with severity-weighted differential omission. Diagnostic and safety beliefs estimate the two loss components, while conformal residual calibration corrects for underestimation from partial evidence. The calibrated risk then guides question selection and stopping by balancing expected risk reduction against interaction cost. Experiments on multiple medical benchmarks show that CaRDI improves accuracy while requiring fewer questions. Further analysis shows that replacing entropy with clinical risk as the planning objective reduces clinical loss by 21.5% and decreases severe omissions. Our code is available at https://anonymous.4open.science/r/CaRDI-A940/.

arXiv:2608.13179v2 Announce Type: replace Abstract: Reinforcement learning with verifiable rewards (RLVR) directly optimizes verifier-defined outcomes for multi-turn tool-use agents, yet its trajectory-level credit assignment conflates heterogeneous per-turn outcomes into a single reward signal. On-policy distillation provides dense per-token supervision but uses a teacher-defined token-level objective and can be prone to gradient concentration collapse. We introduce $\textbf{CrEST}$, a hierarchical credit assignment framework that anchors each sampled token's scalar advantage sign to verified rewards while incorporating dense magnitude signals from a privileged self-teacher. $\textbf{CrEST}$ resolves credit at two levels: turn-segmented verified advantages address inter-turn dilution, while entropy-gated self-teacher modulation refines intra-turn token contributions. Experiments on BFCL V3 and WildToolBench show that $\textbf{CrEST}$ consistently outperforms both RL and distillation baselines across two model scales, with the largest gains on long-trajectory and strict session-level metrics. Our work demonstrates that the teacher's role in policy optimization can be reduced from determining token-level reinforcement or suppression to modulating update magnitudes, enabling dense credit assignment while keeping verifier rewards as the training objective.

arXiv:2608.07943v2 Announce Type: replace Abstract: Multi-page visually rich document understanding (MP-VRDU) systems can fail when evidence is not faithfully represented, not selected, or not used correctly, yet final accuracy alone does not reveal where failures occur. We present an empirical attribution framework that examines these failures through controlled interventions in a unified document question answering pipeline. Experiments on MMLongBench-Doc, with cross-dataset and cross-family checks, show that complementary modalities can reduce sensitivity to conversion errors: images mitigate parser errors, while extracted text reduces dependence on visual resolution. Adding related non-gold pages can both repair incorrect answers and break correct ones, even when all annotated evidence is present, revealing instability beneath aggregate accuracy. Abstention trades unsupported answers for false refusals, while its interaction with reasoning instructions differs across backbones. Qualitative examples and accuracy--cost comparisons connect these findings to guidance on representation, evidence selection, and reasoner choice under deployment constraints.

arXiv:2608.06123v3 Announce Type: replace Abstract: Measuring political bias in large language models (LLMs) remains challenging as it can manifest through subtle differences in framing, argumentation, and legal reasoning that are difficult to capture with a single metric. In this work, we introduce POLI-BIAS, a counterfactual framework for measuring whether LLMs treat legally equivalent conflict scenarios differently depending on the countries involved. POLI-BIAS compares responses to paired prompts in which country identities are systematically swapped across diverse geopolitical relationships, legal violations, and reasoning tasks. Rather than reducing bias to a single judgment, our framework decomposes response disparities into five interpretable dimensions, revealing how and where unequal treatment manifests. Across 5,600 prompt pairs and seven open-weight LLMs from diverse model families and sizes, we find that models treat countries in active conflict or tension less evenly than fictional or uncontested countries, describing, evaluating, and defending equivalent actions differently. Our results thus establish POLI-BIAS as a fine-grained framework for auditing political even-handedness in LLMs.

arXiv:2608.05587v2 Announce Type: replace Abstract: Mobile GUI agents require accurate action reflection for reliable long-horizon execution, assessing whether each action was successfully executed and whether its outcome constitutes a valid step toward the task goal. When evaluated as standalone reflectors under a common prompt, frontier VLMs and specialized GUI models exhibit imbalanced recognition of valid and invalid transitions. We propose Q-Reflector, a standalone reflector based on Qwen3-VL-8B that learns GUI action reflection as supervised structured prediction. Given task context, the executed action, and paired screenshots, it generates pre- and post-conditions specifying the action's prerequisites and intended local effect, then assesses execution success and task validity through a structured rationale and binary verdict. Training combines supervised fine-tuning, teacher-student distillation, and reward- and preference-based refinement. On 1,082 manually verified AndroidWorld transitions, Q-Reflector achieves 82.16% overall accuracy, exceeding all evaluated baselines and showing more balanced performance than GPT-5.2 and GUI-Owl-1.5-8B. On 117 MobileWorld tasks, Q-Reflector achieves task success rates of 46.15% with GUI-Owl-1.5-8B-Instruct and 35.90% with MAI-UI-8B, improving over no external reflection by 17.09 and 14.53 percentage points, respectively. Under the same condition-generation setting, it outperforms the strongest reflection baselines, GPT-5.2 and Claude Haiku 4.5, by 3.42 and 5.98 percentage points, respectively.

arXiv:2607.28236v2 Announce Type: replace Abstract: Pre-trained language models have significantly improved sentence representation learning, yet their embeddings remain sensitive to textual perturbations such as synonym substitution, masking, and word dropout. This work proposes a lightweight Contrastive Denoising Autoencoder (CDAE) that refines pre-trained BERT embeddings by jointly optimizing contrastive and reconstruction objectives to learn perturbation-invariant representations. We evaluate the proposed framework under multiple perturbation strategies and strengths, and compare it against the original BERT embeddings and SimCSE. Experimental results show that CDAE consistently preserves higher embedding similarity under perturbation, with the gap widening as the perturbation strength increases, highlighting perturbation-invariant learning as a promising direction for improving the stability of sentence embeddings. The source code is publicly available at: https://github.com/ComputationIASBS/CDAE

arXiv:2607.22649v2 Announce Type: replace Abstract: Following complex instructions with multiple explicit constraints remains a fundamental challenge for large language models (LLMs). Existing alignment methods, such as DPO, optimize holistic reward signals that often underemphasize strict satisfaction of individual constraints, particularly under out-of-distribution or multi-constraint settings. In this paper, we propose STAIF, a stage-wise optimization framework that decouples the alignment of subjective (soft) constraints from the optimization of objectively verifiable (hard) constraints. Stage 1 applies preference optimization with multiple negative samples to sharpen sensitivity to soft constraints, while Stage 2 applies Reinforcement Learning with Verifiable Rewards (RLVR) to enforce strict compliance with hard constraints. To support this method, we construct STAINSTRUCT, a high-quality bilingual (English, Chinese) dataset of approximately 31,000 complex multi-constraint instructions. Extensive analyses validate the design of STAIF and show state-of-the-art performance on representative benchmarks against strong baselines, as well as genuine generalization.

arXiv:2607.18254v2 Announce Type: replace Abstract: Multi-Level Intermediate Representation (MLIR) underlies modern ML compiler infrastructure (TensorFlow, JAX/StableHLO, PyTorch Inductor, IREE), yet appears only in trace amounts in code-LM pretraining corpora. MLIR is also extensible by design: new dialects ship per application domain, so a fine-tuned model per dialect does not scale. We ask whether inference-time priors derived mechanically from each dialect's Operation Definition Specification (ODS) can substitute for gradient-based adaptation. First, we release four natural-language-to-MLIR benchmarks across three dialects (MLIR-Spec-150, Linalg-Spec-30, StableHLO-Spec-30, StableHLO-Held-Out-200), totaling 410 in-scope NL-to-MLIR pairs, plus a 25-program out-of-grammar stress set, with functional references on 30 Spec prompts, under Apache-2.0 with Gebru datasheets and Croissant 1.1 metadata. Second, we build a three-layer schema-derived constraint stack: a CFG over op signatures (C1), type-domain splits from an ODS-extracted type lattice (C2), and an SSA-scope validator driving five-retry rejection sampling (C3). Porting it to StableHLO required no new constraint-layer code. On structurally dominated dialects, SmolLM2-1.7B with token-level constraints matches or exceeds 15B-34B code LMs running post-hoc rejection at a matched retry budget and 8-25x the per-generation speed: on linalg it reaches 80.0% verify-valid (three runs, n=125), 21 to 44 pp above CodeLlama-34B, Granite-Code-34B, and StarCoder2-15B with non-overlapping CIs. The comparison is between systems, not models: under the baselines' post-hoc protocol, SmolLM2 falls from 52.0% to 24.5% on arith+func. On arith+func and the templated StableHLO-Held-Out-200 the baselines match or beat the SLM; we scope these as non-win cells. We release benchmarks, decoder, every per-prompt generation, and a pinned reproducibility Dockerfile.

arXiv:2607.14240v2 Announce Type: replace Abstract: Current alignment approaches typically focus on emulating human behavior using static representations of human preferences, failing to capture the dynamic, context-dependent nature of real-world human-AI interactions. In this paper, we argue for a shift from static and emulative to interactive and complementary alignment, where preferences emerge through interaction and alignment is defined not by satisfying preferences alone. We first formalize this gap by contrasting existing alignment with a trajectory-level view in which human and model behavior co-evolve over time. Because these interaction dynamics have not been adequately captured within existing ML formulations, we ground this perspective in insights from an interdisciplinary workshop. We draw on lessons from social-science accounts of human-human collaboration and then argue that human-AI systems amplify these dynamics, introducing new asymmetries that make reasoning about uncertainty harder and introduce new coordination challenges. Based on these lessons and new challenges, we conclude by outlining a research agenda for developing AI systems that align with humans in interaction, requiring an interdisciplinary synthesis of machine learning and the social and decision sciences.

arXiv:2607.03162v3 Announce Type: replace Abstract: LLM-powered agents struggle with personalization when users issue raw, underspecified queries. In this setting, agents must infer latent intent, extract preferences from noisy interaction histories, and select among competing alternatives. Existing benchmarks rarely test this capability, as they often rely on user-refined queries or simplified histories. We introduce personalized product search (PPS), a testbed for agentic personalization under raw queries and diverse histories. We construct Agent Personalized Benchmark (APeB) from action logs, pairing underspecified intents with rich histories and user-viewed candidate items. Evaluating state-of-the-art LLMs with multi-step agent workflows, we find that models handle explicit queries well but struggle with early-stage queries requiring intent and preference discovery. Rubric analysis attributes this gap mainly to ineffective history use. A simple history-aware query-refinement pipeline, VQRA, yields consistent gains, highlighting the need for dedicated history-utilization modules in personalized agents.

arXiv:2606.31002v4 Announce Type: replace Abstract: Lean verifies that a generated declaration is well typed, but cannot check whether it states what the user intended. We measure how far compilation overstates faithfulness in statement autoformalization without canonical Lean targets. Our criterion requires Lean compilation and agreement of two LLM judges from different model families. On an independent random audit, it agrees with the human majority on 91.5% of cases; it rarely accepts an unfaithful output but rejects some faithful ones. Across eight systems on 227 university-level statements, the gap between compiling and accepted outputs ranges from 1.3 to 29.5 percentage points: our tool-augmented agent compiles 87.2% of the statements, but only 57.7% are accepted. Human review of all the agent's rejections confirms that 23.8 of its 29.5 points are real translation errors. For the gaps of the seven one-shot systems, a two-stage review by Claude Opus 5.5 with Lean and GPT-6.1 Sol, validated on these human labels, finds that 78-100% are real translation errors for six of them. When GPT-5.2, one of the two judges, grades the translations of two models released after it, GPT-6 Astra and Claude Opus 5.5, it accepts 83% of the statements for each, against 21% for one-shot GPT-5.2. For these two models, the same review classifies only 4 of their 54 rejections (7%) as translation errors, so most of their measured gap is judge errors rather than translation errors. This is consistent with these systems erring less while the judge does not: the fewer errors a system makes, the smaller the share of its rejections that are translation errors. Finally, a 2^3 tool ablation of the agent shows that Lean feedback drives most of its compilation gain, and that nearly half of this gain fails the semantic criterion. Code and data are available at https://github.com/grenadeComing/beyond-compilation.

arXiv:2606.06090v2 Announce Type: replace Abstract: LLM-based agents increasingly tackle long-horizon tasks with interdependent decisions, where each action reshapes future constraints and intermediate errors can cascade. Existing RAG and agent memory systems organize histories by semantic similarity, retrieving content-relevant entries at decision time. We argue that this design mismatches execution-state dependencies: it fragments decision trajectories and mixes valid and erroneous traces, hindering coherent state reconstruction and error isolation. We propose MAGE (Memory as Agent-Guided Exploration), an active execution-state manager that stores interactions in a hierarchical state tree. The agent derives its state from the active root-to-current path, combining subgoal summaries, recent traces, and hints from prior branches. Four coupled operations maintain the tree: Grow records new traces, Compress summarizes completed subgoals, Maintain validates summaries, and Revise restores a target boundary and resumes on a new branch. This design bounds context growth while preserving state integrity and isolating flawed segments from the active path. Experiments on MemoryArena show that MAGE improves the average task success rate by 7.8--20.4 pp over baselines, while reducing token consumption by 55.1%. Code is available at: https://github.com/Komorebi660/MAGE.

arXiv:2606.00991v2 Announce Type: replace Abstract: Transportation systems management and operations (TSMO) increasingly depends on timely interpretation of heterogeneous data, from various sensor streams, incident reports, traveler feedback, and visual observations. Large language models (LLMs), including emerging multi-modal large language models (MM-LLMs), provide a new mechanism for integrating these structured and unstructured inputs into operator-facing decision support. This survey paper reviews LLM- and MM-LLM-based applications in TSMO across three domains: transportation operations & services (supply), mobility & fleet services (demand), and data, modeling & decision support. Using a PRISMA-guided screening process, we synthesize current studies while distinguishing operationally oriented applications from prototype and emerging concepts. We further identify recurring challenges in data heterogeneity, real-time inference, explainability, multi-modal fusion, and governance. Finally, we outline existing gaps and future directions in localized adaptation, edge deployment, benchmarking, and cross-agency collaboration. Overall, LLM-based systems appear most promising as a decision-support layer, with MM-LLMs offering particular value when heterogeneous text, visual, and sensor inputs must be integrated.

arXiv:2606.00729v3 Announce Type: replace Abstract: This paper examines AI sovereignty as national learning capacity, using France as an application. Human-Centered Learning Mechanics (HCLM), proposed for entropy-regulated representation learning (Tran, 2026), provides a conceptual lens for relating information inputs, absorptive capacity and avoidable institutional losses. Inputs include compute, data, talent, research, capital, deployment and institutional experimentation. Potential losses include coordination failures, administrative duplication, energy constraints, regulatory uncertainty and limited organizational uptake. Regulation also supports learning through evaluation, trusted data access, interoperability and safety standards. The argument is that sovereignty depends on converting distributed resources and knowledge into usable, coordinated and responsibly evaluated capability. The analysis connects this account with neural scaling laws, endogenous growth, creative destruction and standard game-theoretic mechanisms. It develops conditional scaling relationships, policy indicators and illustrative scenarios, alongside descriptive OECD, IMF and World Bank evidence. Country observations are kept separate from assumed scenario values, and a fully specified public-goods example shows how decentralized incentives can limit shared provision. The framework offers a research heuristic rather than a literal transfer of neural-network equations to countries or an established measure of sovereignty. Its empirical value requires testing with longitudinal, organizational and cross-country data.

arXiv:2605.29951v2 Announce Type: replace Abstract: A benign image paired with a benign caption can become harmful only in combination, making detection fundamentally a cross-modal reasoning problem. Existing multimodal harm benchmarks evaluate only classification accuracy, leaving unclear whether models identify the cross-modal mechanism that produces harm. We introduce MuPHI (Multimodal Pragmatic Harm Interpretation), a controlled diagnostic benchmark in which harm arises purely from image-text composition, with benign counterfactuals and annotated harm rationales. Using MuPHI, we systematically evaluate where compositional harm reasoning breaks in vision-language models (VLMs). MuPHI reveals that VLMs reliably ground each modality yet fail to bind the evidence into the harm mechanism and their predictions respond weakly to label-determining counterfactual interventions. Moreover, existing harm detection approaches trained on labels substantially improves in-domain (ID) detection but transfers poorly across harm distributions. We therefore propose Compositional Harm Interpretation Learning (CHIL), which rewards evidence identification in each modality, cross-modal binding and verdict consistency in the reasoning chain rather than just predicted label. CHIL improves 81% OOD transfer evaluations by an average of +2.5 macro-F1 and correct counterfactual flip rate by up to +3.8 and +5.4 % points, while also improving overall reasoning quality over label-only optimization. These results suggest that optimizing how a model reasons, not just what it concludes, offers a promising direction for multimodal safety systems that generalize beyond dataset-specific correlations.

arXiv:2605.28360v3 Announce Type: replace Abstract: Automatic prompt optimization (APO) has driven significant gains in LLM-based agentic workflows. However, most existing methods treat each task's prompt as a monolithic, instance-blind string optimized through global edits, producing brittle updates and preventing the reuse of learned sub-behaviors. We propose Prompt Codebook Optimization (PCO), a novel compositional prompt optimization framework that recasts APO as discrete learning over a finite vocabulary of natural-language instincts--atomic, reusable instruction units. PCO organizes prompt-construction knowledge in a discrete codebook and routes each input to a small subset of entries via an LLM-based encoder; a generator composes them into a prompt for the executor; a critic emits a structured verdict that decomposes by attribution into per-variable textual gradients, jointly training the encoder, generator, critic, and codebook under a language-valued min-max objective. The resulting routing is per-instance: different inputs in the same task receive different instinct compositions. Across six benchmarks, PCO improves aggregate performance over zero-shot by +13.50 points on Qwen3-8B and +11.80 points on LLaMA-3.1-8B. Crucially, PCO surpasses GEPA on HotpotQA by +6.34 points (Qwen3-8B) and +4.27 points (LLaMA-3.1-8B), while simultaneously reducing deployed prompt length by up to 14.1$\times$ vs. MIPROv2 and 3.0$\times$ vs. GEPA.

arXiv:2605.25489v2 Announce Type: replace Abstract: Visual analytics (VA) workflows combine data transformation, feature computation, visual representation, and human interpretation, but they are usually described in prose. This hinders comparison of workflows, reuse of proven strategies, and training of novices. We present Artifact--Transform Workflow Language (ATWL), a declarative language for representing the structure and intent of VA workflows. A workflow consists of typed artifacts and of transforms, each with an analytical intent and a human, machine, or hybrid actor. We built a library of 54 workflows extracted from published VA papers with the help of LLMs under human supervision. An exploratory computational comparison of the library workflows shows regularities, such as a regular division of labour between humans and machines and cycles of refinement. We ran experiments in which LLMs designed workflows for new tasks. They show the effects of using the library, the ATWL representation, and both together, and they reveal a limitation in expressing branching that requires an extension of the ATWL syntax. The library is small and partly drawn from our own work, so the findings are preliminary.

arXiv:2605.21993v2 Announce Type: replace Abstract: Ranking systems used for decision support should expose evidence that can be audited independently of the model's ranking scores. We study a closed-roster setting in which a policy jointly emits a Top-$K$ ranking and candidate-linked, span-grounded evidence bundles. Evidence-Coupled Policy Optimization (ECPO) combines listwise utility, certificate validity, and an evidence-cycle reward that removes answer-bearing fields before deterministic recovery. On MAVEN-ERE and RAMS, ECPO improves verifier-audited traceability, rank-blind recovery, and human comparative support relative to matched controls in the tested provenance settings. The gains persist under held-out evaluators, other 8B/14B backbones, and a HotpotQA answer-ranking interface. Ordinary NDCG and runtime remain explicit trade-offs.