Searcharxiv⌕ Search

arXiv subjects

Zihan Dong

Publications and source records attributed to Zihan Dong.

At least 19 recordsLinked to original sources

When the Cheapest Feasible Router Fails: Auditing Deployment Commitments in Cost-Aware LLM Routing

Cost-aware LLM routing often selects the cheapest policy meeting empirical risk and coverage constraints. Yet a router that passes an audit can violate its fixed service contract on new requests, even when candidate policies are constructed independently. We study this deployment-commitment decision with disjoint construction, audit, and test samples. Across 19 RouterBench task families, the cheapest audit-feasible router deploys in 91.82% of opportunities but violates the held-out contract in 32.22% of deployments. We introduce Menu-Aware Routing Commitment (MARC), which qualifies policies using audit evidence and a serving-cost tolerance to retain the apparent winner, deploy a supported alternative, or defer. Its practical policywise variant (MARC-P) reduces observed conditional failure to 0.73% at 85.79% deployment, with higher serving cost; fixed-margin and formal risk-control procedures provide competitive alternatives. Nested audits show how added evidence changes deployable policy sets, and learned semantic routing exhibits the same commitment failure. These results establish deployment commitment as a distinct operational decision connecting router evaluation, service reliability, and inference cost.

cs.LG↗

Agent Benchmarks Stop Ranking Before They Run Out of Tasks

Data-science and ML-engineering agents are evaluated on benchmarks such as MLE-bench, ALE-Bench, SWE-fficiency and AlgoTune, whose tasks yield a quantity (a Kaggle metric, contest score or speedup). When one stops separating the best models, the usual remedy is a harder version. Re-scoring released runs of eight benchmarks, including SWE-bench Verified and Terminal-Bench for breadth, we find much of this saturation lies in the measurement, not the tasks. Most agent benchmarks reduce a run to one pass/fail bit, which fails in two ways. It is too coarse: neighbouring top-ten agents on SWE-bench Verified differ by less than its tests' errors, and correcting them reorders neighbours no more than deleting a random sixth of each agent's passes. It is capped at the reference: ALE-Bench's top ten, separated by contest score, tie when graded against the median human, although nearly half the runs beat that human. On fixed tasks, a higher bar only moves the problem; graded credit fixes it. We derive a clipped log-ratio score with dated anchors and cost as a separate axis. In a Terminal-Bench re-run, pass/fail-tied models differ over tenfold on the quantity the tests discard.

cs.AI↗

Style Over Substance: Content-Invariant Wrappers Flip LLM Safety-Judge Verdicts

Automatic safety judges -- for example Llama Guard, or GPT-4o given a grading prompt -- produce the numbers behind reported jailbreak success rates, defense evaluations, and safety leaderboards. We ask whether these judges grade what a reply contains or how it sounds. We place a fixed piece of text before or after a model reply (a style wrapper) and leave the reply itself unchanged byte-for-byte, so any change in verdict is an error of the judge. Over 600 JailbreakBench replies (300 harmful, 300 refusals), 9 wrappers, and 8 judges, using paired tests and a measured per-judge repeat-disagreement rate (the noise floor), we find that most judges barely move but specific judges have specific, cheap blind spots. A token refusal ("I can't help with that") placed before the unchanged harmful text flips 19.9% of GPT-4o-mini's correct "unsafe" verdicts to "safe" (noise floor 0.5%; 18.2% under majority-of-three re-scoring), yet moves Claude only 0.4%. The deployed Llama Guard 4 guard model is evaded by an "educational course" preface on 12.3% of harmful replies, and the flips persist across repeats, while a second deployed guard (gpt-oss-safeguard-20b) is immune. Changing only the grading prompt to ignore tone cuts the GPT-4o-mini attack roughly tenfold on the same model, so the weakness lives in the judge, not in the content. Pure tone-only decoration that adds no new claim still flips up to 2.1% of verdicts on the vulnerable judges, several times their noise floors. On this artifact set the judges also disagree about which target model is safest, and a bootstrap shows the ranking is unstable before any wrapper is added. We release the data, wrappers, code, and per-verdict labels.

cs.AI↗

When Verifiable Counts Depend on Wording: Auditing Wording Robustness in Instruction Following

Verifiable instruction-following benchmarks often express each constraint through one fixed template. We test whether scores remain stable when the operational requirement is unchanged but its wording varies. We introduce WISE, a matched evaluation suite and reporting protocol instantiated on exact word count, keyword inclusion exactly once, and an inclusive 8--12 word range. Across 100 matched tasks, up to thirteen models from seven providers, and repeated generations scored over the complete visible output, wording alone produces substantial compliance shifts. In an avoidance-family panel, five avoidance and exclusion forms fall below the positive baseline, while constructional controls also shift compliance substantially: in the nine-model control panel, compliance is 54.9% for the original positive form, 48.2% for a longer positive form, 36.7% when the target appears later, and 33.8% for AVOID1. A strict JSON-structure probe shows wording sensitivity beyond counting, with a different direction of effect. Effect sizes, failure directions, weakest forms, and model rankings vary across realizations. Under the most disruptive exclusion form, the top-ranked model changes and 24.1% of strictly ordered model pairs reverse. Human validation further shows that unanimous agreement on an exact-count interpretation can coexist with substantially different model behavior. WISE supplements conventional scores with mean and worst-form compliance, wording gaps, failure profiles, and ranking stability.

cs.CL↗

Count Evidence, Not Sentences: Tempered Evidence Fusion of LLM Judgments for Long-Text Value Measurement

Large language models (LLMs) are increasingly used to measure public value orientations from long social media posts, yet such posts often mix background, quotations, concessions, and only a few stance-bearing sentences. Existing approaches either ask the model to predict a document-level label directly, which can be overconfident, or aggregate sentence-level predictions by majority or soft voting, which treat uncertain and decisive sentences as equally informative. We formulate long-text value measurement as a decision-fusion problem and propose Tempered Evidence Fusion (TEF), a training-free rule that weights each sentence's log-odds by its normalized information gain, as derived from a generalized Bayesian posterior. This makes the fused score nearly vanish for uncertain sentences while preserving the Bayes-optimal weight of decisive evidence. We further introduce Multi-event Insight Network Dimensions (MIND), a benchmark of 8,358 Chinese and English posts spanning five years of public events and six value dimensions. On MIND, TEF outperforms the strongest baseline among Direct, Majority Vote, and Soft Vote by an average of 4.5 accuracy points and 4.6 macro-F1 points across five LLMs and two languages. MIND dataset and code are available at https://github.com/Kzczc/ICASSP2027-TEF.

cs.CL↗

CADWorld: Computer-Use Benchmark for Long-Horizon Computer-Aided Design

Computer-use agents are increasingly evaluated in realistic desktop environments, but existing benchmarks provide limited coverage of professional engineering workflows whose outputs are persistent, structured artifacts. Mechanical computer-aided design (CAD) is a particularly demanding setting: an agent must manipulate geometry and constraints over long interaction horizons while producing a native project whose dimensions, construction structure, and downstream engineering state remain valid. We introduce \textbf{CADWorld}, a benchmark for long-horizon computer use in FreeCAD. CADWorld contains 200 tasks spanning 11 mechanical-CAD workflow categories, including sketching, part modeling, assembly, CAM, FEM, measurement, mesh processing, and technical drawing. Agents operate through screenshots and GUI actions, while success is determined by task-specific executable checks over saved FreeCAD artifacts and auxiliary outputs, covering geometric properties, parametric structure, constraints, manufacturing state, and simulation results. Across seven current agents on the full benchmark, the strongest agent achieves 17.5\% success, compared with an 87.0\% expert reference pass. We find that weaker agents often fail before producing a valid artifact, whereas stronger agents increasingly fail on structural, geometric, and construction-process requirements. CADWorld therefore exposes a gap between general GUI competence and reliable execution of persistent, verifiable engineering workflows. Project accessible at https://cad-world.github.io.

cs.AI↗

Single Canonical Prompts Underestimate LLM Safety's Surface-Form Sensitivity

A benchmark score is a measurement instrument, yet most benchmarks read each item at a single canonical surface form. We ask whether that reading is faithful: when an item's intent is held fixed and only its meaning-preserving surface form varies, does the canonical-form score estimate model behavior well, and how much of any variation is decoding/judge noise rather than signal? We instantiate this in safety, a high-stakes setting with no gold label to average toward. To avoid prior confounds, we pre-author the reformulations (refusal-free, mostly non-LLM: machine back-translation and a Matrix-Language-Frame code-switch generator) so an identical surface form reaches every model, score all responses with one human-anchored, vendor-neutral judge (Claude, kappa = 0.86 vs. human on unsafe compliance, stable across languages, cross-checked by GPT-4o), and verify intent preservation. On 370 seeds x 5 surface forms x 5 models, no single transformation is uniformly most dangerous (6 of 20 per-transformation McNemar tests survive correction, most protective). Yet evaluating only the canonical prompt underestimates unsafe compliance: the union of unsafe outcomes across forms exceeds even the worst single form by 3.3-12.9 pp, with bootstrap 95% CIs excluding zero for all five models, and 5-13% of seeds safe on canonical are unsafe under some reformulation -- above a zero stochasticity floor (canonical resampled five times at temperature 0 gives 0/370 new exposures). The size of this gap is model-dependent (largest on Gemini 2.5 Pro). One form recovers only ~53% of a model's observed unsafe surface and about three reach 85% -- a redundancy characterization of this form set, not of a defined population. A benign control (XSTest) suggests the instability is bidirectional, though the benign and harmful pools are not item-matched. We release the dataset, code, and per-response labels.

cs.CR↗

AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning

Test-time scaling strategies for Large Language Models predominantly rely on either reinforcement learning with sparse outcome rewards or search-based methods guided by static Process Reward Models. However, outcome-based RL often suffers from training instability and sample inefficiency, while static PRMs require expensive step-wise supervision and are susceptible to reward hacking due to distributional shifts. In this paper, we introduce AIRL-S, a unified framework that integrates Adversarial Inverse Reinforcement Learning with Group Relative Policy Optimization. By inferring a dense, step-wise reward model directly from reference trajectories, AIRL-S eliminates the dependency on labeled process data and uses the same learned PRM as both a training signal and a verifier for search-based TTS. Extensive evaluations across eight benchmarks in mathematics, science, and code generation demonstrate that our policy model improves average performance by 9\% over the base model, matching GPT-4o. We further analyze how the AIRL and GRPO objectives complement each other and how the learned PRM transfers across generators and search algorithms, establishing a robust and cost-effective methodology for scaling test-time computation in complex reasoning tasks.

cs.LG↗

Personalization as Inverse Planning: Learning Latent Design Intents for Agentic Slide Generation via Structural Denoising

Slide design requires personalizing both deck themes and page layouts. Yet, current AI agent-based methods struggle with fine-grained, page-level design. Solely relying on prespecified templates or user verbose instructions, they fail to capture latent design intents, leaving Page-level Slide Personalization (PSP) unresolved. To close this gap, this work formulates PSP as an inverse planning problem. We propose to learn a design intent without assuming any knowledge of the specific executing tools (e.g., PowerPoint, Beamer) being used. However, relinquishing control over these tools makes the problem intractable to optimize end-to-end. To overcome this, we propose SPIRE, a principled framework to solve PSP approximately. By intentionally corrupting the visual structures of clean slides, SPIRE creates a verifiable task to denoise the corruption, whereby two agents learn to collaboratively refine executable designs via reinforcement learning (RL). We present a proof that structural denoising is a consistent surrogate for PSP, and that the multi-agent formulation strictly reduces policy gradient variance in RL. Extensive experiments demonstrate the superiority of SPIRE.

cs.AI↗

Hybrid-Policy Self-Editing for Composable Unstructured Knowledge Editing

Large language models (LLMs) achieve remarkable performance across natural language tasks, yet they are trained on static corpora and their knowledge quickly becomes outdated in a fast-changing world. This motivates knowledge editing (KE), which updates specific knowledge in an LLM without changing unrelated others. Recent works move from structured knowledge triples toward unstructured KE (UKE), where the edit is a free-form passage that may state multiple facts at once. Nonetheless, existing editors inject such a passage yet fail to use it: the edited model can recall the passage, but can neither answer atomic questions about its facts nor compose them into multi-hop reasoning. We attribute this missing property, which we term composability, to editors' passive reliance on the fixed passage as the sole learning source. In response, we cast editing as a proactive self-distillation from a privileged in-context state of the same model, which requires no external supervision. We further reveal that due to the novelty of the injected knowledge, the pre-edited model's own rollouts rarely cover it, which limits the effectiveness of pure on-policy distillation. To close this gap, we propose HPSE, which builds a hybrid rollout that steps in to place missing facts onto the student's own trajectory precisely where its coverage fails, while staying on-policy elsewhere. We theoretically analyze HPSE's advantage over pure on-policy distillation, and empirically establish its plug-and-play improvements across four LLM backbones and two KE editors under various scenarios.

cs.CL↗

How Benchmarks Mis-Score Computer-Use Agents

Computer-use agents (CUA) are being deployed to browse the web and operate desktop software, yet their benchmark scores are still commonly produced by brittle scripted oracles. A score is the output of a pipeline in which tasks can be stale, trajectories can omit decisive visual evidence, evaluators can reject valid alternatives, and aggregate reports can hide the cause of failure. We organize these problems into a reliability framework spanning task construction, trajectory observation, scoring, and reporting. We then audit 150 public failure-scored trajectories from five web, enterprise-workflow, and desktop-control benchmarks, find that 15.3\% of FAIL verdicts are wrong: 10.7\% are evaluator false negatives and 4.7\% are broken tasks. For genuine failures, a three-tier diagnostic taxonomy shows that verification/feedback and planning failures dominate execution/grounding errors, while a single scalar success rate can not explain. We connect these findings to newer long-horizon CUA benchmarks and derive stage-specific design rules for CUA evaluation.

cs.AI↗

Why Are GUI Agents Correct but Late? Decode on the Decision-Time Critical Path, Tested with Pre-Compiled Policy Trees

Computer-use agents often fail on transient GUI events because they produce the correct action only after the relevant window has already closed. We identify the main cause as expensive autoregressive decoding on the decision-time critical path. We propose Adaptive Anticipatory Policy Trees (AAPT), which eliminates this delay without modifying the underlying model. During idle screen periods, the same frozen multimodal model constructs a bounded conditional policy tree with observable guards, pre-authorized actions, and branch-specific deadlines. The tree is sized to cover the model's own decoding latency. When an event occurs, a lightweight observer matches change-gated frames to a prepared branch and immediately executes the corresponding action without generating new text. In paired trials with pre-registered endpoints and exact McNemar tests, AAPT improves the success rate from 0.50 to 0.79 within a contested decision window ($p=1.8\times10^{-3}$), while producing no incorrect actions. Both open-loop and predict-and-replan baselines achieve zero success because they still decode during execution. A preparation-time sweep shows that the gain emerges where the latency-based tree-sizing rule predicts, and ablations reveal three key requirements: fast observer decoding, valid tree planning, and accurate branch routing. A pre-registered oracle probe rejects our initial hypothesis and instead points to branch routing as the causal bottleneck. We further reproduce the effect on an independent general-purpose multimodal model over 126 paired trials ($p=4.9\times10^{-13}$). On an external benchmark, AAPT matches the overall performance of a reactive baseline, although the two methods exhibit complementary strengths. Together, these results suggest that AAPT performs best when candidate actions can be enumerated in advance, whereas reactive execution remains stronger when they cannot.

cs.LG↗

Evaluating LLMs When They Do Not Know the Answer: Statistical Evaluation of Mathematical Reasoning via Comparative Signals

Evaluating mathematical reasoning in LLMs is constrained by limited benchmark sizes and inherent model stochasticity, yielding high-variance accuracy estimates and unstable rankings across platforms. On difficult problems, an LLM may fail to produce a correct final answer, yet still provide reliable pairwise comparison signals indicating which of two candidate solutions is better. We leverage this observation to design a statistically efficient evaluation framework that combines standard labeled outcomes with pairwise comparison signals obtained by having models judge auxiliary reasoning chains. Treating these comparison signals as control variates, we develop a semiparametric estimator based on the efficient influence function (EIF) for the setting where auxiliary reasoning chains are observed. This yields a one-step estimator that achieves the semiparametric efficiency bound, guarantees strict variance reduction over naive sample averaging, and admits asymptotic normality for principled uncertainty quantification. Across simulations, our one-step estimator substantially improves ranking accuracy, with gains increasing as model output noise grows. Experiments on GPQA Diamond, AIME 2025, and GSM8K further demonstrate more precise performance estimation and more reliable model rankings, especially in small-sample regimes where conventional evaluation is pretty unstable.

cs.LG↗

RIMS: Preference Optimization via Smoothed Multi-pair Aggregation for Small-Scale LLM Retrieval-Augmented Generation

Small-scale language models (SLMs) are attractive for retrieval-augmented generation (RAG) in resource-constrained settings, but their limited capacity makes them highly sensitive to noisy or spurious retrieved evidence. Existing preference-based methods such as RoseRAG select only the hardest single preference pair via hard argmin/argmax, discarding the remaining signal; others treat multiple pairs as independent binary comparisons, resulting in low data utilization. We propose RIMS, a three-stage preference optimization framework comprising (1) synthetic chain-of-thought preference data generation via rejection sampling using the target SLM itself without relying on proprietary models, (2) a differentiable soft aggregation mechanism that replaces hard selection with a smooth operator, preserving gradient signal from all preference pairs while retaining the discriminative structure of margin-aware selection, and (3) preference optimization with the smoothed objective applied to multiple alignment algorithms. We theoretically show that the smoothed approximation admits a controllable error bound and that smooth aggregation yields provably tighter gradient alignment to the oracle objective than hard selection. Experiments on four multi-hop question answering benchmarks show that our approach outperforms state-of-the-art baselines across multiple SLM backbones, achieving consistent gains in Exact Match and F1 under noisy retrieval conditions. Our implementation is available at https://github.com/tptrix29/RIMS.

cs.CL↗

Conformal Path Reasoning: Trustworthy Knowledge Graph Question Answering via Path-Level Calibration

Knowledge Graph Question Answering (KGQA) offers grounded, interpretable reasoning, but existing methods often fail to provide reliable coverage guarantees over retrieved answers. While Conformal Prediction (CP) offers a principled framework for producing prediction sets with statistical guarantees, prior conformal KGQA methods suffer from two critical pitfalls: violated coverage guarantees due to invalid calibration, and weak score discriminability that yields excessively large prediction sets. We propose Conformal Path Reasoning (CPR), a novel trustworthy KGQA framework built on two key innovations. First, query-level conformal calibration over path-level scores preserves exchangeability to ensure valid coverage guarantees. Second, we introduce the Residual Conformal Value Network (RCVNet), a lightweight module trained via PUCT-guided exploration to learn discriminative path-level nonconformity scores. Extensive experiments show that CPR significantly improves the Empirical Coverage Rate by 45% while reducing prediction set size by 52% on average over conformal baselines across benchmark datasets, highlighting its effectiveness for reliable conformal reasoning over knowledge graphs.

cs.CL↗

Decompose Sparsely Where You Should, Absorb Densely Where You Should No

Sparse autoencoders (SAEs) are typically trained to reconstruct the \textbf{entire} residual stream through a sparse dictionary, implicitly assuming that all activation content is amenable to sparse, monosemantic decomposition. We question this assumption and hypothesize that activations contain a low-rank, dense component that is computationally important to the model yet inherently unsuitable for sparse representation, which serves as a major source of the persistent dense latents widely observed in trained SAEs. To test this, we add a small rank-$r$ linear bottleneck in parallel with standard SAEs (BatchTopK and Matryoshka), allowing dense structure to be absorbed before sparse reconstruction. On Gemma-2-2B layer 12, a rank-24 bottleneck reduces dense latent count by up to 84\% while improving sparse probing and targeted probe perturbation on both architectures at matched sparsity. The absorbed component is (i) \textbf{structurally identifiable} as the top principal components and outlier dimensions; (ii) \textbf{causally necessary}, with removing it raising next-token cross-entropy by 7.5$\times$, far exceeding the 2.8$\times$ from removing the geometrically near-identical top-24 PCA directions; and (iii) \textbf{redundantly encoded by sparse dictionaries}, with ablating 787 maximally aligned sparse features raising cross-entropy by only 2.9$\times$ and ablating 2,048 topic-aligned features leaving MMLU topic classification virtually unchanged, whereas removing the scaffold drops it from 98.7\% to chance. Together, our findings identify a compact, semantically informative and causally important component of residual stream activations (which we term a \textbf{computational scaffold}) that standard sparse dictionaries represent inefficiently, suggesting that the scope of sparsity-based interpretability methods warrants careful re-examination.

cs.LG↗

Evidence Over Plans: Online Trajectory Verification for Skill Distillation

Agent skills can remarkably improve task success rates by using human-written procedural documents, but their quality is difficult to assess without environment-grounded verification. Existing skill generation methods heavily rely on preference logs rather than direct environment interaction, often yielding negligible or even degraded gains. We identify that it is a fundamental timing bottleneck: robust skills should be posterior-based, distilled from empirical environment interaction rather than prior plans. In this study, we introduce the Posterior Distillation Index (PDI), a trajectory-level metric that quantifies how well a distilled skill is grounded in the task-environment evidence. To operationalize PDI, we present SPARK (Structured Pipelines for Autonomous Runnable tasKs and sKill generation) for preserving task execution evidence towards full trajectory-level analysis. SPARK generates environment-verified trajectories used to compute PDI, and it applies PDI as an online diagnostic and intervention signal to ensure posterior skill formation. Across 86 runnable tasks, SPARK-generated skills consistently surpass no-skill baselines and outperform human-written skills on student models (inference cost up to 1,000x cheaper than teacher models). These findings show that PDI-guided distillation produces efficient and transferable skills grounded in the task-environment interaction. We release our code at https://github.com/EtaYang10th/spark-skills .

cs.AI↗

RUBRIC-ARROW: Alternating Pointwise Rubric Reward Modeling for LLM Post-training in Non-verifiable Domains

Pointwise reward modeling offers critical signals for LLM post-training, yet struggles with absolute scoring in subjective, non-verifiable settings. Rubric-based methods address this by decomposing evaluation into explicit criteria, but existing approaches typically depend on frontier LLMs and suffer from ties caused by hard Boolean aggregation. We present RUBRIC-ARROW, an alternating framework that jointly trains a rubric generator and a rubric-conditioned judge, with its RL stage using only pairwise preference data. Our method couples a probability-based scoring rule that reduces ties with phase-specific preference-based rewards and an alternating GRPO scheme that together train the pointwise evaluator. Extensive experiments show that RUBRIC-ARROW achieves competitive reward-modeling accuracy and yields consistent gains for downstream policy post-training.

cs.LG↗