Artificial Intelligence · Cross-listed from physics.chem-ph
RxnOptBench: Benchmarking LLMs for Reaction-Condition Optimization in Organic Methodology
Lingli Ge, Yubin Wang, Junyuan Gao et al. (+9 more)
Chemical reaction-condition optimization – choosing the catalyst, ligand, solvent, reagent, temperature, time, and atmosphere that jointly maximize yield and stereoselectivity – is a central, judgement-laden subtask of organic methodology research that large
language models are increasingly expected to support. Yet existing chemistry
benchmarks evaluate reaction-class labelling, retrosynthesis, or SMILES manipulation, and do not ask models to read a real condition-screening table and pick the best set. We introduce RxnOptBench, a
benchmark whose every option and precedent is a real wet-lab entry mined from the optimization tables of organic-methodology papers published in 2025, graded by a continuous relative score derived from a declared headline utility that combines reported yield with enantiomeric excess (ee), diastereomeric ratio (dr), and regioisomeric ratio (rr), and equipped with a paired precedents-vs-no-precedents design that isolates in-context use of literature evidence from parametric memorization. Across nine frontier LLMs and three Chemistry LLMs, even the best models leave substantial headroom: chemistry-specialized models fall to the random-baseline floor on multi-axis selection, while open-weight models have closed most of the gap to proprietary frontier models. We release the final human-reviewed
benchmark test set and evaluation code.
Read the full abstract ▾
All authors Lingli Ge, Yubin Wang, Junyuan Gao, Jiahe Song, Jiaxing Sun, Boyu Zhu, Haote Yang, Jingchao Wang, Lixin Ma, Jiang Wu, Yuqiang Li, Conghui He
Artificial Intelligence · Cross-listed from cs.AR
Hardware-Native Joint Sparse-Quantization for Trillion-Scale Mixture-of-Experts
Kwanhee Lee, Namhoon Lee and Dan Alistarh
Mixture-of-Experts (MoE) architectures allow frontier
language models to scale to trillions of parameters, but their deployment is constrained by massive
memory footprints and
memory-bandwidth limitations. Although modern accelerators provide Sparse Tensor Cores (SpTCs) that reduce weight storage and increase throughput through low-precision semi-structured sparsity, exploiting them for MoEs remains challenging because of substantial model-quality degradation and the lack of grouped sparse GEMM primitives. We present an end-to-end hardware-software co-design framework that compresses expert weights into hardware-native, low-precision sparse representations and accelerates their execution on SpTCs. Algorithmically, our framework relaxes discrete semi-structured support selection through continuous reparameterization, enabling differentiable joint optimization with quantized weights under a router-weighted reconstruction objective and scalable expert-parallel compression. Systemically, we develop a custom grouped sparse GEMM kernel tailored to low-precision sparse MoE inference on SpTCs. Across MoE models ranging from 30 billion to one trillion parameters, our framework improves state-of-the-art joint sparse-quantization accuracy by up to 4.35 percentage points while preserving 96.09% of the original model's performance. On NVIDIA B200 GPUs, our kernel outperforms the vendor baseline by up to $1.65\times$, increasing serving throughput by $1.18\times$ and reducing end-to-end latency by up to $4.03\times$. These results establish hardware-software co-design as a practical path toward scalable and efficient MoE deployment.
Read the full abstract ▾
Artificial Intelligence
Fault-Tolerant Budget Conservation in Distributed Multi-Agent Delegation
Genliang Zhu and Chu Wang
Resource limits are becoming an authorization boundary for AI
agents that delegate work across concurrent and failure-prone workers. Parent-child allocation constraints, affine objects, and distributed escrow do not by themselves prevent overspend when replies are lost, effects complete after timeout, messages repeat, branches partition, or DAG joins alias one lineage. We formalize fault-tolerant budget conservation for distributed multi-
agent delegation. Budgets are quantized resource vectors represented by exclusive escrow credits that move through a delegation DAG. Before dispatch, a branch converts credit into an operation reservation bound to lineage, epoch, normalized effect, maximum charge, receiver, and idempotency key. It persists a signed dispatch permit with quarantine; the gateway verifies that permit before first acceptance. Uncertain effects remain charged until authenticated settlement, a fenced authoritative no-effect proof, or permanent retirement. We prove ownership partition, ledger and effect conservation, descendant non-amplification, at-most-once settlement, late-completion safety, and partition confinement under explicit mediation, durability, authentication, normalization, and gateway assumptions. An indistinguishability result shows that partition-local availability requires exclusive preallocation. Bounded TLA+ checking, an independent JavaScript explorer, and crash-injected two-process SQLite experiments exercise the declared scope and detect timeout-refund and historical-certificate-validation mutants. The mechanism preserves the issued budget bound across the evaluated crash, retry, duplicate, partition, join, and late-completion schedules.
Read the full abstract ▾
Artificial Intelligence · Cross-listed from cs.CR
Authorization for Self-Modifying AI Agent Populations: Conserving Authority across Replacement, Forking, and Rollback
Genliang Zhu and Chu Wang
Self-modifying AI
agents can replace, fork, and roll back identity-bearing software while descendants remain executable. Per-successor authorization does not constrain the resulting population: siblings may duplicate quotas, combine permissions, survive ancestor cuts, or overlap predecessors during promotion. We define authorization succession, which conserves authority across the active frontier of a single-parent generation forest. Our external protocol binds each generation to a manifest, root, unique parent, complete lineage, and fresh population sequence. Separate invariants bound root-lifetime consumption and current population exposure. A staged reservation freezes predecessor residual authority during replacement, while a partitioning fork validates the complete child family. Each commit atomically fences the predecessor and activates successors. Ancestor cuts invalidate dependent descendants; rollback creates a fresh generation without restoring spent authority; and a new root requires an independent grant. Under complete mediation, authenticated records, sound effect abstraction, durable monotone state, and complete lineage accounting, we prove population-safe succession, fork conservation, revocation closure, atomic handoff, rollback non-reminting, and exclusion of self-certification. An executable evaluation covers 32 registered decisions through direct-call and mailbox mappings (64/64 replays; 28 allows, 36 denies). An independent checker accepts all 64 original traces and rejects 28/28 semantic mutants; 12/12 profile invariants, 16/16 crash cuts, and 32/32 contender schedules pass. Two external adapters reproduce all 32 decisions around measured OurArk and Darwin Godel Machine mutations, including fresh-process restart, atomic succession, and predecessor rejection. The results establish authorization succession for registered protected effects.
Read the full abstract ▾
Artificial Intelligence · Cross-listed from cs.CV
LEGO-OPD: Factorized Teacher Composition for Multimodal On-Policy Distillation
Jaeyun Shin, Hangeol Chang and Jong Chul Ye
Multimodal on-policy
distillation (OPD) aims to improve visual grounding while preserving the strong
reasoning capabilities of
language models. Recent multi-teacher approaches combine LLM and VLM teachers to provide complementary supervision. However, directly using a VLM's full predictive distribution entangles its visual grounding signal with its own
language prior, preventing the grounding information from being transferred independently. Conversely, increasing the strength of visual supervision can improve perception but may overemphasize visual evidence and degrade
language reasoning. To address this trade-off, we introduce LEGO-OPD, which selectively composes factors from a
Language Expert and a Grounding expert into One teacher distribution for
multimodal OPD. Under a generalized
Bayesian formulation, the
language expert provides a prior over candidate
tokens, while the grounding expert contributes a visual likelihood that updates this prior, rather than transferring its complete predictive distribution. This factorized composition allows
language reasoning and visual grounding to be controlled independently. We further introduce adaptive calibration to determine how strongly the visual likelihood should update the
language prior at each decoding prefix. Specifically, LEGO-OPD uses the grounding expert's image-induced prediction shift as a prefix-dependent reference, preventing both insufficient and excessive visual supervision. Experiments with Qwen3 models show that LEGO-OPD consistently outperforms the evaluated single- and multi-teacher OPD baselines on both
multimodal and text-only
reasoning tasks. Moreover, it improves the initial student's visual perception while preserving text-only
reasoning.
Read the full abstract ▾
Artificial Intelligence · Cross-listed from cs.LG
The Weakest Link: Distilling LLM Reasoning with Worst-Case Constrained Reinforcement Learning
Matthieu Zimmer, Xiaotong Ji, Tu Nguyen et al. (+1 more)
Distilling the
reasoning capabilities of large
language models (LLMs) into smaller students is a central challenge for efficient deployment. Current approaches face a fundamental tension: optimizing purely for verifiable task rewards (e.g., via GRPO) leads to reward hacking, where students arrive at correct final answers through flawed intermediate logic, while regularizing with soft divergence penalties against a teacher (e.g., KL-based
distillation) dilutes task performance and, critically, allows the student to compensate for severe logical violations at one step with high teacher agreement at others. We argue that this averaging is fundamentally misaligned with the nature of
reasoning: a chain-of-thought is only as valid as its weakest link. Motivated by this observation, we formulate
reasoning distillation as a constrained reinforcement learning problem in which the task reward is maximized subject to a worst-case constraint on the teacher log-likelihood along every prefix of the trajectory. To avoid the prohibitive cost of dual Lagrangian solvers and the test-time teacher dependence of state-augmented methods such as Saute, we derive an unaugmented constrained MDP whose reward transformation preserves the hard-constraint semantics, admits a low-variance policy gradient decomposition into single-step and long-term terms, and provably satisfies the worst-case constraint almost surely in the penalty limit. Through extensive experiments on mathematical
reasoning and code generation tasks, we demonstrate that our method significantly expands the accuracy-fidelity Pareto front. By matching the high Final Answer Correctness of pure RL and drastically reducing teacher constraint violations, we ultimately achieve the highest rigorous
Reasoning Success Rate across all evaluated settings.
Read the full abstract ▾
All authors Matthieu Zimmer, Xiaotong Ji, Tu Nguyen, Haitham Bou-Ammar
Artificial Intelligence
Mathematical Transfer in LLMs Follows Reasoning Approach More Than Topic
Sajad Goudarzi, Samaneh Zamanifard, Seyed Amin Seyed Haeri et al. (+2 more)
When selecting mathematical training data for LLMs, a natural organizing principle is topic: probability examples for probability targets. An alternative is
reasoning approach: worked solutions that share a solution method with the target, even when the mathematical domain differs. We ask which relation produces greater transfer after fine-tuning. We evaluate two counterbalanced $2\times2$ designs: probability and combinatorics crossed with invariant
reasoning and double counting (2,000 problems), and number theory and geometry crossed with complement and pigeonhole
reasoning (800 problems). In each design, every
cell serves as the held-out target in turn: same-approach (SA) sources share the target's method but change the topic, while same-topic (ST) sources share the topic but change the method. Every source appears once in each role, so additive source-quality effects cancel from the equally weighted aggregate contrast. Across five base models and three training seeds per design, SA outperforms ST in all 40 seed-pooled model–target comparisons. Model-level advantages range from 8.2 to 16.2 percentage points in the primary design (mean: 10.8) and from 12.0 to 16.0 in the second design (mean: 14.3); all ten model-level 95% confidence intervals exclude zero. In both designs, ST sources are more similar to targets under
embedding and lexical measures, so the SA advantage runs opposite to the measured ordering of statement-level resemblance. These findings identify
reasoning approach as a more effective matching criterion than topic for mathematical transfer across the evaluated topic–approach combinations.
Read the full abstract ▾
All authors Sajad Goudarzi, Samaneh Zamanifard, Seyed Amin Seyed Haeri, Moloud Nasiri, Hamed Rahimian
Artificial Intelligence
ContractRL: Shielded Group-Relative Policy Optimization for Auditable Tool-Call Repair
Miaobo Hu, Shuhao Hu, Xiaobo Guo et al. (+4 more)
Structured tool calls often fail after only a small number of fields violate a schema or an execution contract. Regenerating the complete object enlarges the action surface and makes repeated repair difficult to audit. We introduce ContractRL, a contract-constrained sequential repair protocol that models verifier-guided JSON repair as a bounded decision process. At each step the policy observes the candidate, typed verifier feedback, JSON Pointer, immutable repair history, and remaining budget; a contract-derived action mask filters malformed or prohibited RFC-6902 operations before a deterministic validator performs the transition. We specify a contract-constrained group-relative objective for patch, retry, and abstention decisions while keeping canonical targets and semantic labels outside the online state until trace freeze. Under identical verifier information, ContractRL attains 0.9362 semantic success with 34.4 generated
tokens, compared with 0.9076 and 44.9
tokens for Patch-SFT and 0.9148 and 137.2
tokens for full regeneration over 192 cases per seed and five seeds. Policy optimization improves semantic success from 0.9186 for supervised ContractRL to 0.9375. A separate three-seed paired evaluation against Patch-SFT yields a semantic difference of +0.0396 (95% CI $[+0.0137,+0.0662], p=0.0039$). Feedback, action-mask, budget, and schema-shift analyses connect these gains to localized correction, while
adversarial and multi-turn evaluations characterize the remaining failure modes.
Read the full abstract ▾
All authors Miaobo Hu, Shuhao Hu, Xiaobo Guo, Xin Wang, Bokun Wang, Yina Sa, Daren Zha
Artificial Intelligence · Cross-listed from cs.CR
Actions with Receipts: Jointly Binding Claims, Evidence, and Execution for Replayable Tool-Agent Auditing
Miaobo Hu, Shuhao Hu, Xiaobo Guo et al. (+4 more)
Tool-using
agents can expose citations and execution logs while leaving a critical association unaudited: whether the claim shown to a user is the claim emitted by the committed execution and supported by the cited source. A valid citation and a valid trace can therefore remain individually well formed while being transplanted across claims, actions, runs, or source versions. We introduce a claim-anchored execution contract that jointly binds the emitted claim, its exact source span, the ordered execution prefix that produced it, and the source version and access state observed by that execution. Each receipt contains an emission anchor that deterministically locates the claim inside a committed answer or claim-bearing action, together with source identifiers, offsets, hashes, quotes, and a domain-separated execution commitment. A deterministic integrity verifier reconstructs these bindings before semantic or task labels are joined. We separate this integrity plane from a pluggable support plane, so structural validity is not used as a proxy for entailment. The contract exposes seven independently testable properties: claim-emission binding, source binding, ordered-execution binding, oracle separation, persisted-object replay, execution-rerun consistency, and version/access binding. Across 1,280 cross-object attacks, the joint contract detects 1,275 substitutions (0.9961). Removing a targeted property reduces its attack-detection rate to 0.0156-0.0625. On an independently adjudicated 384-pair split, the conflict-aware support guard reaches F1 0.8865 and false acceptance 0.0729; on unseen failure families, these rates are 0.8679 and 0.0938.
Read the full abstract ▾
All authors Miaobo Hu, Shuhao Hu, Xiaobo Guo, Xin Wang, Bokun Wang, Yina Sa, Daren Zha
Artificial Intelligence · Cross-listed from cs.RO
DriftOPD: Sequence-Level Reverse-KL Distillation for One-Step VLA Policies
Youngjun Jun, Kyumin Choi, Youngmin Kim et al. (+4 more)
Vision-
Language-Action (VLA) models increasingly rely on action experts that generate short action chunks under receding-horizon control. While chunk-level training is convenient across
robot embodiments, it optimizes local action likelihood without explicitly accounting for long-horizon task success. Sequence-level reinforcement learning can address this limitation, but typically requires policy rollouts and closed-loop interaction, which are costly for real-
robot manipulation. We introduce DriftOPD, a teacher-free, rollout-free framework for sequence-level on-policy
distillation of continuous VLA action experts. We show that the sequence-level reverse Kullback-Leibler (KL) divergence decomposes into a chunk-level reverse-KL term and a future-potential term that captures the long-horizon effect of the current action. DriftOPD optimizes these two terms using a one-step drifting objective and a Q-function critic learned from offline demonstrations, respectively, enabling sequence-level optimization with only offline data and one-step action generation. Across multiple VLA architectures in simulation and real-world manipulation, DriftOPD generally outperforms existing one-step
distillation baselines while achieving task success performance comparable to multi-step teacher policies. These results demonstrate that long-horizon behavior can be effectively distilled into one-step VLA action experts without online interaction or a separate teacher.
Read the full abstract ▾
All authors Youngjun Jun, Kyumin Choi, Youngmin Kim, Seonghyun Jin, Sunwoo Park, Jangho Park, Jong Chul Ye
Artificial Intelligence · Cross-listed from cs.CL
DuplexSpeechBench-Document Grounding: Benchmarking Document Grounding and Hallucinations in Voice Agents
Puneet Mathur, Nedim Lipka, Zeyu Jin et al. (+1 more)
Voice
agents enable low-latency, natural interaction, yet their ability to faithfully ground responses in external documents remains underexplored. We introduce DuplexSpeechBench-Document Grounding (DSB-DG), a
benchmark for evaluating document grounding in voice
agents across five professional domains. DSB-DG targets three failure modes: Context Saturation, which measures grounding under increasing document length; Grounding Decay, which measures retention of document facts across multi-turn dialogue; and Proactive Grounding, which evaluates whether context re-injection mitigates conversational drift. The
benchmark contains 1,636 adversarially verified QA pairs from 50 documents covering five professional domains, and supports fully automatic evaluation of grounding accuracy,
hallucination, and response latency. Across systems spanning cascaded, proprietary full-duplex and real-time, and open-weight
speech2
speech architectures, we find substantial differences in effective grounding capacity. While cascaded pipeline (ASR-LLM-TTS) achieves the highest grounding accuracy, Gemini-Live and GPT-Realtime closely trail behind. Open-weight systems exhibit distinct failure modes, most notably an abrupt context-capacity collapse and multi-turn grounding decay. More broadly, grounding fidelity degrades with context and conversational load, and failures frequently manifest as unsupported generations rather than abstention. We show that contextual grounding as a key unresolved challenge for reliable full-duplex voice
agents.
Read the full abstract ▾
All authors Puneet Mathur, Nedim Lipka, Zeyu Jin, Dinesh Manocha
Artificial Intelligence · Cross-listed from cs.CV
Beyond Pixel Reconstruction: Retrieval-Guided Glyph-Aware Restoration for Low-Resource Manchu Historical Documents
Ting Huang, Dongdong Wang, Mingqiu Liang et al. (+1 more)
Historical Manchu documents preserve invaluable linguistic and cultural heritage, yet their digitization is hindered by severe degradations and the scarcity of paired training data. Existing document restoration methods primarily optimize pixel-level reconstruction, which can produce visually plausible results while failing to preserve the structural identity of Manchu glyphs. To address this limitation, we propose a
retrieval-guided glyph-aware restoration framework that goes beyond pixel reconstruction by explicitly incorporating glyph-level structural knowledge. Our method retrieves relevant glyph exemplars to provide structural guidance during restoration and integrates this information into the reconstruction process, improving the recovery of degraded character structures under low-resource conditions. Extensive experiments on Manchu historical documents demonstrate that the proposed approach improves both image restoration quality and glyph-level fidelity compared with existing restoration methods. These results highlight the importance of incorporating character-aware structural priors for reliable restoration of low-resource historical documents.
Read the full abstract ▾
All authors Ting Huang, Dongdong Wang, Mingqiu Liang, Siyang Lu
Artificial Intelligence
Predictive Credit: Measuring What Scientific Explanations Add to Experimental Forecasts
Jingjie Ning, Xueqi Li, Yibo Kong et al. (+1 more)
Research
agents explain planned experiments. We measure predictive credit with paired forecasts sharing an intervention, forecaster, and outcome while varying description, matched explanation, and donor context. Five checks track commitment, delivery, predictive gain,
alignment, and known-signal uptake. Across 336 prospective states in controlled learning, 12 Tox21 endpoints, and 24 OpenML tasks, v5's frozen credit decision was inconclusive. Tox21's preregistered ROC AUC interval-score harm test was unmet ($D-M=-.0026$, 95 percent interval [$-.0174$, .0104]); OpenML's joint formation, point-equivalence, and repeatability rule was unmet. Matched point-accuracy gains over description remained unconfirmed, and Tox21/OpenML seed-donor intervals spanned zero. Under requested DeepSeek V4 Pro, matched and donor cards reduced secondary Tox21 drift by 64.5 and 59.1 percent. A DeepSeek V4 Flash replay raised matched point MAE from .01823 to .02020 and missed matched-donor interval-score equivalence. OpenML full-card assignment widened nominal 80 percent intervals by 21 percent, with 49.3 percent coverage versus 51.4 percent for description and content in 66/144 cards. Direct-text Flash delivered all 144 notes without detectable matched point-accuracy gain. A researcher-authored mechanism positive control lowered point MAE by 2.60 percentage points versus description. The protocol measures predictive credit for research-
agent benchmarks and scientific forecasting; natural-explanation credit remained unconfirmed at the tested donor resolutions.
Read the full abstract ▾
All authors Jingjie Ning, Xueqi Li, Yibo Kong, Dongting Li
Artificial Intelligence
Rules to Tools: Executable Checks for LLM Agents in Scientific Computing
Jingjie Ning, Guojiang Zhao, Chen Xu et al. (+4 more)
Scientific coding
agents receive equations, boundary conditions, and output requirements in writing, then must assess the programs they revise. Rules to Tools (R2T) supplies prepared executable checks of public scientific requirements. Matched SciCode repair groups share written checks, starting programs, model, and budgets; the tool group receives a callable implementation. Across two task-ID cohorts, complete repair is 26/30 with text and 29/30 with the prepared checks. Three task IDs favor tools, one favors text, and eleven tie. The eight-ID cohort scores 13/16 versus 15/16, with a task-cluster bootstrap 95% interval of [-12.5, 43.75] percentage points for the difference. The larger shared-definition SciCode cohort ties at 13/24 per group. Five development-exposed tasks with alternate starting programs score 3/10 versus 7/10. The tool group favors tasks 17, 77, and 11; initial checks flag task 17 and report no violation for tasks 77 and 11. Task 37 favors text and has no initial reported violation. A fresh source-through-Python arm also reaches 15/16, matching the dedicated command's aggregate. In a matched PDE comparison, detailed text scores 23/24 and checks score 24/24, with 31.2% lower reported model output for checks.
Agent-side output savings vary by cohort, while public CPU use rises in both task-ID cohorts. These results measure task-dependent repair outcomes and
agent-side costs with prepared checks.
Read the full abstract ▾
All authors Jingjie Ning, Guojiang Zhao, Chen Xu, Shanshan Zhong, Xiaochuan Li, Ji Zeng, Guolin Ke
Artificial Intelligence · Cross-listed from cs.LG
Provable Test-Time Scaling for Beam Search in LLM Reasoning
Qijia He, Yu Huang, Yuan Cheng et al. (+2 more)
Beam-search-based test-time methods provide an effective way to improve large
language model (LLM) performance on long-horizon generation by
pruning invalid
reasoning paths early, leading to significantly improved
reasoning efficiency and more favorable test-time cost scaling. Despite strong empirical success, the theoretical understanding of beam search remains limited. In this paper, we study the test-time compute guarantee of the commonly used beam search framework that uses the model's internal log-likelihood for intermediate scoring, while relying on an external reward model only after a complete response is generated. We first establish a lower bound for vanilla beam search, showing that at least $\Omega(C^\star(x)^2)$ samples are required for the optimal response to survive, where $C^\star(x)$ is the
token-level coverage coefficient for
prompt $x$. This motivates our modified confidence-filtered beam search (CF-Beam), which reduces the sufficient coverage dependence from quadratic to nearly linear under prefix competitiveness, for fixed horizon, gap, and target accuracy. We then show that the regret of CF-Beam is upper-bounded by the probability of rare failure events and the reward estimation error scaled by a path-level coverage coefficient, where the rare-failure term vanishes as per-step sampling increases. Our results highlight a fundamental advantage of beam search over sequence-level inference methods such as Best-of-N and Best-of-Majority. While the guarantees of these approaches typically involve coverage coefficients that grow exponentially with the horizon $L$, CF-Beam controls the dominant search-induced term through a
token-level coverage coefficient that scales polynomially with $L$. Our numerical experiments further confirm that beam search is more robust on hard instances and under increasing
reasoning horizons.
Read the full abstract ▾
All authors Qijia He, Yu Huang, Yuan Cheng, Yuxin Chen, Yingbin Liang
Artificial Intelligence
Where Scientific Search Agents Fail: Decision-Checkpoint Auditing of Exposure and Inspection Attempts
Hongmin Li and Wanli Zhao
Final-answer accuracy does not reveal whether a scientific-search
agent failed to encounter a target paper, attempt to inspect it, or return an accepted answer after inspection. We introduce decision checkpoints that record observations and tool actions without
benchmark labels during inference, then join target identities and evaluator labels to assign outcome categories from recorded events. Across five conditions on 540 answerable AutoResearchBench Deep questions in a fixed, target-enriched environment, keyword search achieves 24.6% accuracy, compared with 17.8% for raw search. The keyword condition has fewer incorrect answers with neither target exposure nor inspection, but more incorrect answers after the target is exposed and left uninspected. Compared with keyword search, read-first has 27.4% more recorded evidence-search calls. Target inspection attempts occur on 199 questions under read-first and 191 under keyword search; both conditions achieve 24.6% accuracy. The checkpoint protocol makes these question-level differences explicit, distinguishing target exposure and inspection from aggregate accuracy and total tool use.
Read the full abstract ▾
Artificial Intelligence · Cross-listed from cs.LG
Understanding Off- vs On-Policy Distillation: A Tale of Distinct Training Objectives
Qiwei Di, Xuheng Li, Kaixuan Ji et al. (+3 more)
On-policy
distillation (OPD) learns from teacher feedback on student-generated responses and has shown promise in reducing forgetting relative to supervised fine-tuning (SFT). However, its benefits and fragility remain incompletely understood. We study sequential
distillation from multiple teachers, where the student minimizes its average divergence from the teachers. Forward Kullback–Leibler (KL) divergence yields a weighted arithmetic mixture, while reverse KL yields a normalized weighted
geometric aggregate. We develop
algorithms that learn these targets under off-policy and on-policy feedback, respectively, establishing logarithmic regret bounds in the tabular setting and extending the analysis to function approximation. By analyzing these aggregation targets, we identify mechanisms that help explain both the benefits and fragility of OPD. Relative to forward KL, reverse KL can better retain a confident expert's preferences under uninformative feedback, but is more sensitive to teachers that assign very low probabilities to correct responses. Its
token-level conditionals also reveal a dependence on continuation distributions that can favor incorrect prefixes over long horizons.
Read the full abstract ▾
All authors Qiwei Di, Xuheng Li, Kaixuan Ji, Chenggong Zhang, Heyang Zhao, Quanquan Gu
Artificial Intelligence · Cross-listed from cs.MA
CollabFlow: Recursive Self-Improvement of Agent Collaboration
Xiao Huang, Mingda Zhang, Junming Zhang et al. (+5 more)
Recursive self-improvement (RSI) lets a system improve from its own outcomes; in LLM-based multi-
agent systems,
Agents refine one another within a task, and outcomes improve how they collaborate across tasks. However, existing multi-
agent collaboration leaves this loop open: collaboration is pre-defined at the operator level,
topology-only learning keeps verbatim exchange that propagates errors, and reward maximization on a system's own outcomes concentrates on a few teams. To address these challenges, we propose CollabFlow, an RSI system of Learned
Agent Collaboration: a trainable Collab-Director constructs teams of complete
Agents, a frozen executor runs them, and each round's outcomes retrain the director. Within each round, the edges of a collaboration graph carry protocols of Evidence-Conditioned Communication: a receiver adopts a differing answer only when the sender's evidence is stronger by a margin, so the director learns who communicates and how. Across rounds, we further propose Collaborative Trajectory Balance (CTB), a flow-based objective that credits each team once across its construction orders and targets a reward-proportional distribution over teams, so several good teams stay in play. We also bound how far this self-generated target moves between rounds, which shrinks as records accumulate. On twelve
datasets, CollabFlow outperforms all baselines and keeps improving across rounds. Code is available at https://anonymous.4open.science/r/CollabFlow-631E.
Read the full abstract ▾
All authors Xiao Huang, Mingda Zhang, Junming Zhang, Qiang Huang, Hanwen Zhang, Yue Dai, Zijia Wang, Xiaoying Tang
Artificial Intelligence
EvoSteer: Online Self-Evolving Graph Orchestration via Reference-Anchored Credit Assignment
Mingda Zhang, Hanwen Zhang, Qiang Huang et al. (+5 more)
In recent years, LLM-based multi-
agent systems have been widely applied to orchestrate tool-using
agents into executable communication graphs. However, existing self-evolving orchestration still faces key challenges, including post-hoc
evolution that revises the team only after the trajectory ends, credit
diffusion that gives every action the same terminal advantage under confounded baselines, and skill admission that is uncalibrated and never retired. To address these challenges, we propose EvoSteer, a new paradigm of Online Self-Evolving Graph Orchestration – the orchestrator builds a running team and repairs its plausible but failing steps from execution features and a learned value estimate. To support this paradigm, we introduce Anchored Trajectory Balance (AnchorTB), a regression-style flow-matching loss that assigns each orchestration action a coefficient by balancing subtrajectories against a frozen reference. Built on the learned flow, we further propose Validated Skill Admission, in which a candidate skill is tried before promotion and promoted only if paired evidence passes a sequential test under a shared nominal testing budget. Moreover, AnchorTB combines measured task-level reference reward statistics with prefix-dependent corrections. Experimental results on twelve
datasets show that EvoSteer significantly outperforms baselines across question answering, mathematical
reasoning, code generation, and interactive decision making. Our code is available at https://github.com/beita6969/evosteer.
Read the full abstract ▾
All authors Mingda Zhang, Hanwen Zhang, Qiang Huang, Zijia Wang, Pengfei Guo, Yuchen Zhang, Jionghao Zhu, Xiaoying Tang
Artificial Intelligence · Cross-listed from cs.CL
Breaking Babel: A Self-Evolving Multi-Agent System for Long-Form Subtitle Translation
Haibo Jin, Xinjie Li, Najmeh Sadoughi et al. (+4 more)
Long-form subtitle translation requires
reasoning over discourse and cultural context spanning episodes or entire series, while maintaining consistent terminology and style. Existing single-LLM methods are largely sentence-level, and multi-
agent systems often use static workflows that do not adapt to scene complexity or production context. We propose SMART, a Self-evolving Multi-
Agent system for long-foRm subtitle Translation. During test-time training, SMART builds persistent series-level
memory and translates a subset of sentences through a dynamic router and Mixture-of-
Agents layer with tools for terminology verification, subtitle constraint validation, and contextual
retrieval. A judge-refiner loop scores candidates and uses textual critiques to update
agent prompts and routing policies without retraining the underlying LLMs. During test-time inference, the evolved configuration translates the remaining series. We also introduce Subtitle Arena, covering 14 genres, 2–198 episodes per series, production years 1959–2023, and 15 target locales, together with SubMQM, a subtitle-adapted MQM framework with seven dimensions and 19 error categories. SMART achieves the best overall MQM score in all 15 Subtitle Arena directions, reducing average penalty by 6.9% over the strongest competing
agent system. On a public
benchmark, MuSC, SMART obtains the best model result across all 4
language pairs. SMART also achieves the best result in human evaluation with an overall score of 4.50/5.
Read the full abstract ▾
All authors Haibo Jin, Xinjie Li, Najmeh Sadoughi, Yang Liu, Yibo Wang, Zhu Liu, Yuzong Liu
Artificial Intelligence · Cross-listed from stat.ML
Bandits with Multiple Optimal Arms: Minimax Regret and Non-Adaptivity
Kaixuan Ji, Qiwei Di, Qingyue Zhao et al. (+2 more)
We study multi-armed bandits (MAB) with multiple optimal arms, motivated by the fact that many practical decision making problems admit multiple correct answers. For $K$-armed bandits with $A$ optimal arms, we first provide a sharper analysis of previous sub-sampling
algorithms (De Heide et al., 2021; Zhu and Nowak, 2020), establishing a $\tilde{O}\Big(\frac{K-A}{\sqrt{KA}}\sqrt{T} \Big)$ minimax regret, where $T$ is the total number of interactions and $\tilde O(\cdot)$ drops all constant and logarithmic factors, improving the previous $\tilde{O}(\sqrt{KT/A})$ regret. We then provide a matching lower bound up to logarithmic factors, indicating that our established rate is nearly minimax-optimal. We further show that the knowledge of $A$ up to $\tilde{O}(1)$ factors is necessary to achieve near-optimal regret, as near-optimal
algorithms for one number of optimal arms must incur substantially larger regret than optimal regret for a smaller number. Overall, our results provide a comprehensive minimax characterization of $K$-armed bandits with $A$ over the entire range of $1 \leq A \leq K-1$.
Read the full abstract ▾
All authors Kaixuan Ji, Qiwei Di, Qingyue Zhao, Heyang Zhao, Quanquan Gu
Artificial Intelligence · Cross-listed from cs.RO
TERRA: Terrain-Aware Reconstruction, Retargeting and Control for Musculoskeletal Locomotion
Merkourios Simos, Chengkun Li, Bianca Ziliotto et al. (+1 more)
Recent advances in musculoskeletal modeling and reinforcement learning have enabled muscle-actuated
agents to reproduce increasingly complex human motions. Yet these capabilities remain largely confined to flat ground, in part because motion
datasets rarely include aligned terrain geometry and because retargeting terrain interactions to complex musculoskeletal bodies is challenging. We present TERRA, an end-to-end pipeline for terrain-aware retargeting and control of musculoskeletal locomotion. From kinematic trajectories alone, TERRA combines terrain priors, estimated contacts, and negative free-space evidence to recover task-relevant support geometry. TERRA further considers anatomical, tendon-continuity, and contact constraints during retargeting. Using the resulting motion-terrain pairs from five
datasets, we successfully train a single muscle-actuated control policy on 9.4 hours of diverse locomotion. Across reconstruction, retargeting, and held-out tracking
benchmarks, TERRA improves terrain accuracy, sharply reduces anatomical and interaction violations, and achieves the highest observed completion rate over supported terrain families. Overall, TERRA provides a practical route from scene-less motion data to muscle-actuated locomotion over diverse non-flat terrain. Project website: https://cnai.epfl.ch/terra/
Read the full abstract ▾
All authors Merkourios Simos, Chengkun Li, Bianca Ziliotto, Alexander Mathis
Artificial Intelligence
AgBench: Agentic AI Benchmarks for Personal AI Devices
Yizhou Han, Di Wu, Dhananjay Saikumar et al. (+1 more)
Agentic AI systems increasingly rely on cloud-hosted large
language models for planning, tool use, and iterative execution, raising concerns about API cost and data exposure. Advances in personal AI devices enable
agents to execute locally, but limited resources on device may affect task success and performance. Existing
benchmarks are inadequate for systematically characterizing these trade-offs across devices, workloads, and deployment architectures. We present AgBench, a
benchmark suite and open artifacts for reproducible evaluation of agentic AI on personal devices. Using AgBench, we evaluate local, hybrid, and cloud execution across agentic workloads, examining task success, latency, cloud API cost, and data exposure. Our results, drawn from over 162.07 million data points, show that personal AI devices can complete many
agent tasks locally, but local-only execution generally has lower task success and longer completion times than cloud-only execution, especially as concurrency increases. Local-only execution eliminates cloud model API costs and sensitive-information exposure to cloud
agents. Hybrid execution can improve task success, but its cloud cost and data exposure depend on how
agents divide work and share information. No single architecture performs best across task success, goodput, cloud cost, and data exposure; deployment choices should reflect the intended workload and device capabilities. AgBench is available at https://anonymous.4open.science/r/AgBench-2777.
Read the full abstract ▾
All authors Yizhou Han, Di Wu, Dhananjay Saikumar, Blesson Varghese
Artificial Intelligence · Cross-listed from cs.LG
Alignment via Training Against Probes Without Losing Monitorability
Lena Libon, Alexander Panfilov, Ben Rank et al. (+3 more)
Models are usually aligned based on their observed outputs, using demonstrations, preference data, or reward signals. These objectives reward responses that look aligned. More capable models may learn to satisfy them without internalizing the intended behavior, for example by faking compliance during training. Such superficial compliance could be harder when the objective is defined on model internals rather than outputs. Therefore, we study probe-guided fine-tuning, using probes that detect undesired properties in model activations as a direct training signal. We evaluate linear and non-linear probes with different numbers of probes per layer across two
alignment objectives: harmlessness and honesty. We find that training against probes that do not update during training is an easily exploitable objective, while continuously updated probes substantially reduce harmfulness and improve honesty while preserving utility. Probe-guided fine-tuning achieves better safety-utility trade-offs than DPO and inference-time steering, while being substantially more robust against jailbreak and abliteration attacks. Moreover, the concepts stay linearly encoded after fine-tuning, meaning oversight is not lost by our method. Training against probes thus offers a way to shape what models represent rather than only what they output, which may become increasingly important as models get better at making their outputs look aligned.
Read the full abstract ▾
All authors Lena Libon, Alexander Panfilov, Ben Rank, Xin Chen, Jonas Geiping, Maksym Andriushchenko
Artificial Intelligence
ChartRevise: A Dataset and Evaluation Protocol for Exact Chart Editing via Code
Jiaxiang Tang, Yi Zhou, Chad DeLuca et al. (+5 more)
Chart editing requires cross-modal edit grounding, realizing a requested visual change in the code that draws it, with necessary related updates and without altering unrelated content. Existing
benchmarks emphasize either code executability or chart quality, but their metrics do not clearly distinguish request completion from missed coupled updates and gratuitous changes. We introduce ChartRevise, a structured
dataset and evaluation protocol for exact program-grounded chart editing. For
dataset construction, we build on the grammar of graphics to systematically cover chart-editing operations, using source-program checks to verify their applicability across chart types and libraries. To improve edit exactness, our pipeline checks individual requirements and guides repair or exclusion when they are unmet. The resulting
dataset contains 92,438 records covering 344 edit types across 20 chart types and three plotting libraries. For evaluation, our reference-free protocol separately measures atomic requirement completion, identifies gratuitous changes, and detects missed coupled updates. These checks are combined with successful execution and rendering to determine exact-edit success. Across five models and four external
benchmarks, fine-tuning yields relative gains of 16% in mean requirement recall and 22% in mean exact-edit rate.
Read the full abstract ▾
All authors Jiaxiang Tang, Yi Zhou, Chad DeLuca, Rogerio Feris, Ahmed Khalil Omran, Zhi-Li Zhang, Pengyuan Li, Ali Anwar
Artificial Intelligence
Component-Aware Feedback for Self-Evolving Programs
Ethan Lin, Jinming Nian and Yi Fang
LLM-guided evolutionary search can discover complex programs, but existing methods mostly only save candidate programs and fitness scores while discarding which component edits produced which fitness metric changes. Existing methods force the mutator LLM to infer the effect of prior edits from cluttered histories, making program search slow and unstable. This is especially true for locally servable LLMs to evolve multi-component systems. We introduce component-aware feedback, which compares each evaluated program with its parent, identifies the components that changed, and logs them with the associated metric differences into an attribution
memory that later mutations read. The
memory keeps each change in two reference frames, local against the parent it came from and global against the seed program, which shows both the immediate effect of a change and the cumulative progress made since the seed. We study this on LLM reranking, a multi-objective optimization problem where a multi-stage pipeline must balance quality against serving cost. Across twelve Bright
datasets, our method reaches the strongest baseline's final quality after a median of one third of the search budget and ends 7.2% higher in held-out nDCG@10, and under a cost-aware objective it finds pipelines that are on average more accurate while using 11% fewer
tokens per query, showing component-aware feedback to be a promising direction for more efficient self-evolving systems.
Read the full abstract ▾
Artificial Intelligence · Cross-listed from cs.CV
Template-Search Domain Adaptation via Multi-Stage Feature Alignment for Cross-Modal Object Tracking
Fereshteh Aghaee Meibodi, Amir Mehdi Soufi Enayati, Shadi Alijani et al. (+1 more)
Visual object tracking typically assumes that the initial template and subsequent search frames share the same sensing modality. In practice, sensor availability or operation may change over time, creating a substantial representation gap between template and search frames. Unlike conventional multi-modal tracking where paired modalities are simultaneously available, cross-modal tracking requires localization when template and search frames originate from different active modalities. Accordingly, we introduce TSDA-Track, a Template-Search Domain Adaptation framework to reduce modality discrepancy during training. We investigate two feature
alignment strategies. Pre-AFA TSDA-Track applies
adversarial alignment before
transformer's template-search interaction to suppress modality-specific bias. Enc-CFA TSDA-Track applies contrastive
alignment to encoder representations after interaction to strengthen target-level cross-modal correspondence. Both variants retain a shared inference pipeline without modality-specific branches. Experiments on LasHeR, and zero-shot evaluations on RGBT234 and GTOT under multiple cross-modal protocols demonstrate improvements over representative state-of-the-art trackers. For instance, under the modality-switch protocol on RGBT234, Pre-AFA TSDA-Track achieves an SR/PR of 43.2/56.0, compared with 36.8/50.0 for ToMP-101 baseline. In addition, a study on Anti-
UAV-024 further verifies the applicability of TSDA-Track to aerial tracking. Our study highlights the effectiveness of feature
alignment domain adaptation for cross-modal tracking.
Read the full abstract ▾
All authors Fereshteh Aghaee Meibodi, Amir Mehdi Soufi Enayati, Shadi Alijani, Homayoun Najjaran
Artificial Intelligence · Cross-listed from cs.LG
Interpretable but Fragile? Robustness of Concept Bottlenecks under Geometric-Semantic Perturbations
Hanwei Zhang, Tianma Hu, Gaojie Jin et al. (+2 more)
Concept Bottleneck Models (CBMs) are designed to provide interpretable intermediate representations, yet how such bottlenecks affect robustness remains unclear, with existing studies reporting mixed and sometimes contradictory findings. We argue that these discrepancies arise from conflating different robustness notions and perturbation regimes, rather than from fundamental disagreements about CBMs themselves. To disentangle these factors, we introduce a generator-based evaluation framework that enables controlled comparisons between standard classifiers and CBMs under two distinct perturbation types: continuous
geometric perturbations in latent space and discrete semantic interventions in concept space. Within this framework, we evaluate robustness both empirically, via prediction and concept-level sensitivity metrics, and certifiably, using randomized smoothing in latent and concept spaces. Across experiments, we reconcile previously conflicting findings by clarifying when, and in what sense, concept bottlenecks do or do not improve robustness. By further analyzing robustness under varying task conditions, including class semantic similarity and concept vocabulary size, we show that interpretability does not inherently confer robustness. Instead, concept bottlenecks shift where and how sensitivity manifests, revealing a nuanced interpretability robustness trade off that depends critically on the perturbation regime and task structure. Together, our results show that interpretability and robustness are distinct objectives: interpretable intermediate representations do not uniformly improve robustness, but instead redistribute sensitivity across perturbation spaces and model families.
Read the full abstract ▾
All authors Hanwei Zhang, Tianma Hu, Gaojie Jin, Xu Cheng, Ronghui Mu
Artificial Intelligence
When Scientific Contradictions Are Lost in Translation
Tal Zeevi, Trey W. Jensen and Maxwell Strome
Two scientific findings can disagree without contradicting each other. Determining whether they conflict requires knowing whether they describe comparable measurements. We study how
language models behave at this decision point. In a controlled task, we generate an unsatisfiable XOR constraint system and translate its constraints into scientific reports from different laboratories. One assignment satisfies more constraints, while another satisfies fewer but better matches expected biology. This creates a simple dilemma: does the model choose the assignment that best fits the constraints, or the one that better matches biological expectations? When the constraints are stated directly, GPT-5.6 Sol and Claude Opus 5 recover the best-supported assignment in 90% and 96% of cases, respectively. In scientific prose, however, the models behave differently. Claude Opus 5 often prefers the biologically expected assignment. Removing that biological preference increases recovery of the better-supported assignment from 27% to 79% (p<.001); recovery reaches 92% when the same Biology-favored record is accompanied by a formalization request and an explicit paired-design cue (p<.001). GPT-5.6 Sol is less sensitive, with neither corresponding change reaching statistical significance. These results suggest that reliable scientific verification depends not only on formal
reasoning, but also on how models decide which findings should be compared and what relations they imply.
Read the full abstract ▾
Artificial Intelligence · Cross-listed from cs.LG
Differentiable Structure Learning for Cyclic Linear Gaussian Models with Latent Confounders
Sadegh Khorasani, Ali Najar, Saber Salehkaleybar et al. (+1 more)
We study causal structure learning from observational data in linear Gaussian structural causal models in the presence of directed cycles and an unknown number of exogenous latent confounders, bounded by a given maximum. We derive the covariance of the observed variables and introduce marginal quasi-equivalence, which characterizes when different causal models share a full-dimensional subset of the observational distributions they can generate. We formulate structure learning as minimization of the Gaussian negative log-likelihood with a logarithmically scaled complexity penalty that counts directed edges and latent variables. For a fixed number of observed variables and a fixed upper bound on latent variables, we establish consistency of global score minimizers up to marginal quasi-equivalence under algebraic faithfulness, structural minimality, and model-overlap assumptions. We parameterize the inclusion of directed edges and candidate latent variables using Bernoulli gates, whose continuous probabilities are optimized jointly with the structural coefficients. Averaging the penalized negative log-likelihood over these gates yields an objective with a closed-form differentiable complexity penalty. We prove that this expected objective has the same global infimum as the corresponding discrete structure-learning objective. Experimental results show that our approach achieves lower recovery error than previous methods in several experimental settings.
Read the full abstract ▾
All authors Sadegh Khorasani, Ali Najar, Saber Salehkaleybar, Negar Kiyavash