Searcharxiv⌕ Search

arXiv subjects

Qishi Zhan

Publications and source records attributed to Qishi Zhan.

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↗

Beyond Marginal Risk: ESG and Multidimensional Fragility in Equity Markets

Do ESG characteristics provide financial resilience beyond conventional risk, and is that inference stable to the financial state and ESG score being measured? We study a point-in-time panel of nonfinancial S&P 500 constituents from January 2019 through November 2025. Once ex ante return, volatility, and liquidity exposures are modeled explicitly, the broad downside-resilience pattern attenuates and none of the three primary marginal ESG $\times$ Down tests survives Holm adjustment under either conventional or wild-cluster-bootstrap inference. A post-primary exact decomposition then shows that the negative Return+RVOL coefficient is almost entirely allocated to observations that also breach the CHL trading-cost tail, while the mutually exclusive Return+RVOL-only component is essentially zero. Finally, a paired standardized comparison on an identical sample shows that MSCI's downstream Industry Adjusted Score and upstream Weighted Average Key Issue Score imply materially different fitted associations in adverse market states: at a 5% market decline, the difference is about one probability percentage point per score standard deviation. Sector-time, leave- one-month-out, and observed-change timing diagnostics preserve most of this contrast, although the result remains exploratory and inference is sensitive to multiplicity and market-state support. The evidence therefore favors a narrower interpretation of ESG resilience: the apparent association weakens once conventional risk is modeled, while the remaining relation depends on both the adverse state and the MSCI score field used to represent ESG.

q-fin.MF↗

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↗

ValuSpec: Plug-and-Play Candidate Valuation before Target Verification for Tree-Based Speculative Decoding

Tree-based speculative decoding accelerates large language model inference by verifying multiple candidate branches in parallel. However, in practical serving scenarios, similar requests repeatedly induce overlapping candidate branches. While retrieval-based methods mitigate this by reusing historical drafts, the resulting candidate trees often contain many low-quality branches as the candidate tree grows, which reduces verification efficiency and wastes target-side computation. To address this, we propose ValuSpec, a plug-and-play candidate valuation mechanism that filters low-quality branches from a supplied candidate tree before expensive target-side verification. ValuSpec introduces transfer-based supervision to align filtering with the target model's behavior. A filtering threshold is estimated from the target model's greedy trajectories on a calibration split, without any parameter training. The target model then verifies the filtered tree under the output-identical greedy rule, so accepted tokens are guaranteed to match standard autoregressive decoding exactly. Experimental results on HumanEval, GSM8K, Dolly, and SpecBench show that, in our main evaluation, ValuSpec achieves up to 2.38x end-to-end speedup in retrieval-based tree settings. Furthermore, integrating ValuSpec into the EAGLE-3 dynamic tree framework yields up to 2.78x end-to-end speedup without modifying its tree construction strategy, demonstrating that ValuSpec serves as a generally applicable and plug-and-play candidate valuation module.

cs.AI↗

Why On-Policy Distillation Sometimes Fails: Vanishing Learning Signals

On-policy distillation (OPD) enables effective capability transfer between language models, yet the mechanisms underlying its failures are not fully understood. Across code generation and mathematical reasoning, OPD with larger-scale teachers exhibits early loss plateaus, with an average final loss reduction of 25.1% after 200 updates, compared with 96.2% for self-RL teachers, obtained by further reinforcement learning (RL) training of the initial student. To understand this difference, we analyze OPD as an idealized continuous-time dynamical system in the small-learning-rate limit. Our training-log diagnostics associate these plateaus with an early decline in a gradient-based learning-signal proxy while substantial loss remains; these measurements do not establish why the underlying gradient weakens. We further prove a local recovery guarantee for teachers sufficiently close to the initial student in a shared parameterization under regularity conditions, offering a conditional explanation for the success of self-RL teachers in our experiments. Across runs with and without loss plateaus, we observe small relative parameter changes (0.025-0.098%) and high similarity between the student's representations before and after OPD (linear CKA $>0.98$ across layers). These observations suggest that limited representation adaptation may contribute to learning-signal collapse, a hypothesis that remains to be tested. Code is available at https://github.com/leizhao7/opd-learning-signals.

cs.LG↗

Dual-Pathway Circuits of Object Hallucination in Vision-Language Models

Vision-language models (VLMs) have demonstrated remarkable capabilities in bridging visual perception and natural language understanding, enabling a wide range of multimodal reasoning tasks. However, they often produce object hallucinations, describing content absent from the input image, which limits their reliability and interpretability. To address this limitation, we propose Dual-Pathway Circuit Analysis, a framework that identifies and characterizes hallucination-related circuits in VLMs for mechanistic understanding and causal probing. We first apply activation patching across five architecturally diverse VLMs to identify a visual grounding pathway that supports correct predictions and a hallucination pathway that drives erroneous outputs. We then introduce Conditional Pathway Analysis (CPA) to characterize pathway-level interactions, revealing that grounding components remain strongly redundant in both correct and hallucinating samples but undergo a consistent polarity flip, shifting from supporting the ground truth on correct samples to aligning with the hallucinated answer on erroneous ones. We further perform targeted suppression of hallucination-pathway components, showing that scaling these components reduces object hallucination by up to 76% with minimal accuracy cost, and validate that the same circuit selectively transfers to relational but not attribute hallucination. Evaluations on POPE-adversarial and AMBER show that the identified circuits are consistent across architectures, support causal intervention, and transfer selectively across hallucination types. Code is available at https://github.com/jiaxin26/DualPath-VLM.

cs.CV↗

How Order-Sensitive Are LLMs? OrderProbe for Deterministic Structural Reconstruction

Large language models (LLMs) excel at semantic understanding, yet their ability to reconstruct internal structure from scrambled inputs remains underexplored. Sentence-level restoration is difficult to evaluate automatically because scrambled sentences often admit multiple valid reorderings. We introduce OrderProbe, a deterministic benchmark for structural reconstruction using fixed four-character expressions in Chinese, Japanese, and Korean, which have a unique canonical order and thus support exact-match scoring. We further propose a diagnostic framework that evaluates models beyond recovery accuracy, including Semantic Accuracy, Logical Validity, Structural Consistency, Robustness, and Information Density. Experiments on twelve widely used LLMs show that structural reconstruction remains difficult even for frontier systems: zero-shot recovery frequently falls below 35%. We also observe a consistent gap between meaning-oriented generation and exact structural reconstruction, suggesting that structural robustness is not an automatic byproduct of semantic competence.

cs.CL↗

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↗

Behavior Pack Optimization for Video MLLM Post-Training

Video multimodal large language models (MLLMs) keep climbing video question answering benchmarks, yet shuffling the frames, masking the segment that supports the answer, or occluding the target object barely changes their predictions. The accuracy rests on appearance and language priors, not on the temporal evidence the question asks for. We trace this to the unit of post-training: rewards are computed on a single response to the original clip, so the model is never asked to behave consistently across views. We propose Behavior Pack Optimization (BPO), which replaces the single response with a behavior pack of outputs across counterfactual views chosen by question type, scored jointly. The pack reward asks for stability when the intervention is irrelevant, sensitivity when key evidence is removed, and abstention when no evidence remains. To keep this objective stable at small pack sizes, BPO uses an anchor-relative advantage: the response on the original view serves as a per-prompt reference instead of a group mean over mixed views. On TempCompass, MVBench, and NExT-QA, BPO improves the macro accuracy of Qwen2.5-VL-7B-Instruct by 4.7 pp, the temporal-hard subset by 7.8 pp, and abstention F1 by 20.0 pp over a budget-matched vanilla GRPO baseline from the same SFT checkpoint. The gains transfer to Video-MME, LongVideoBench, and to LLaVA-Video-7B; ablations confirm they follow the view sets, not the rollout count. We hope this pack-level perspective offers a useful starting point for the video MLLM and multimodal post-training community as the field moves toward evidence-grounded video reasoning.

cs.CV↗

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↗

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↗

Discontinuous Prior-Mode Sections and the Geometry of Ambiguity in Intrinsic Image Decomposition

The viral 2015 photograph known as "The Dress" divides observers into two camps because it is ambiguous: the same image colors can be explained either as a blue-black surface under one illuminant or as a white-gold surface under another. We propose a geometric account in which the ambiguity arises from a singularity in intrinsic image decomposition, the inverse problem of separating an observed image into reflectance and illumination. Our central claim is that the prior-mode section, i.e. the prior-preferred decomposition, switches across an ambiguity boundary in image space, and that any smooth learned model can only approximate this discontinuous switch by forming a thin transition layer. This predicts two observable signatures, where $Δ$ is the jump between branches of the prior-mode section and $λ$ is the regularization strength: for inverse decomposers, an albedo Jacobian scaling as $|Δ|/\sqrtλ$; and for forward encoders, the Fernet curvature that blows up on a scale of $1/\sqrtλ$. On CGIntrinsics ($N=1998$ images, $n=2\times 10^7$ pixels), the color-temperature albedo Jacobian of Careaga DPT has partial Spearman correlation $r=0.41$ with dense ground-truth albedo error, compared with $r=0.087$ and $r=0.021$ for brightness and saturation controls. On "The Dress", CLIP ViT-L/14 exhibits a latent curvature peak of $κ=73.03$ at $6473\,\mathrm{K}$, one sampled step from D65 daylight, while a control dress image peaks at $κ=34.75$ with no comparable feature near D65. The same characteristic appears across architectures (U-Net inverter, diffusion inverter, ViT encoder) and datasets (rendered indoor scenes, web photograph), each measured with the observable appropriate to its model class.

math.DG↗

Stress Amplified Resilience: ESG and Joint Fragility in Equity Markets

Market stress rarely harms investors through one channel alone. Losses, volatility spikes, and deteriorating tradability often arrive together. We examine whether ESG is associated with lower exposure to clustered fragility in equity markets. Using monthly data on S&P 500 constituents from 2014 to 2025, we study downside returns, volatility, illiquidity, and a cofragility state that captures their joint occurrence within the same firm month. The evidence supports a stress-amplified resilience interpretation rather than an unconditional ESG return premium. In the return channel, the ESG association is concentrated in the extreme downside tail during stress months. In the volatility channel, higher ESG is associated with smaller risk spikes when aggregate conditions are weak. In the illiquidity channel, the association is more persistent, suggesting a liquidity-quality component whose relevance increases when market-wide trading conditions deteriorate. The central evidence comes from the joint analysis: a one-standard-deviation increase in ESG lowers the stress-period probability of severe cofragility by 0.92 percentage points, about 9% relative to the baseline. Double Machine Learning shows a similar negative ESG association after flexible adjustment for observable firm characteristics. Pillar evidence suggests stronger baseline resilience for Environmental scores and clearer stress amplification for Social scores. Overall, the findings characterize ESG as a multi-channel fragility signal for tail-risk monitoring, stress analysis, and pillar-level ESG assessment.

q-fin.MF↗

BudgetDraft: Acceptance-Aware Multi-View Training for Sparse-KV Speculative Decoding

Speculative decoding speeds up autoregressive decoding by using a drafter to propose multiple tokens that a verifier validates in parallel. In resource-constrained deployments, the drafter uses a sparse KV cache to limit peak GPU memory and end-to-end latency under a fixed KV budget, while the verifier keeps a full KV cache. Mid-to-long context inference (4K--16K context length) is common in real applications. However, naive sparse/full speculative decoding suffers from the sparse/full mismatch as context length grows, causing the acceptance rate to drop quickly. We propose BudgetDraft, a multi-view sparse training method for sparse drafting in mid-to-long inference. The drafter is exposed to multiple sampled KV budgets during training and learns to align each sparse view with one shared full-cache teacher target. BudgetDraft combines an acceptance-aware loss on a full-cache branch with a multi-view loss on a sparse-cache branch, producing a single budget-robust drafter that recovers acceptance across sparsity levels without extra inference-time components. Experimental results on PG-19, LongBench, and LWM show that BudgetDraft achieves up to 6.55x, 4.46x, and 2.10x end-to-end speedup vs AR at 4K, 8K, and 16K context lengths, while keeping the inference pipeline memory-friendly.

cs.LG↗

Disentangling Language Roles in Multilingual LLM Task Execution

Multilingual LLMs are increasingly used when instruction, source content, and required response languages do not coincide. Existing benchmarks have expanded multilingual instruction-following evaluation, but they rarely isolate these three roles within a fully crossed design. We introduce MTM-Bench, a controlled benchmark for language-conditioned task execution in which each instance is defined by a triplet \((L_{\text{instr}}, L_{\text{content}}, L_{\text{resp}})\). Across English, Spanish, and Chinese, MTM-Bench enumerates all 27 triplets and contains 2{,}430 instances per model across semantic reversal, final-state extraction, and language purity with update realization. We evaluate 20 frontier and open-weight LLMs using decomposed metrics for semantic correctness, target-language adherence, constraint satisfaction, contamination ratio, and joint success, with scoring validated by a targeted human audit. The fully crossed design reveals that degradation is organized by the role a language occupies in the task structure, not merely by mismatch count. The response-language role is the dominant axis of variation, and a single response-slot mismatch accounts for most degradation. The response-only and full-mismatch comparison suggests that mismatch count is not a monotonic predictor of difficulty, with model-level ordering varying across systems. Task families fail through distinct channels, showing that semantic correctness alone does not capture reliable multilingual task execution.

cs.CL↗

Relative Repairability: A Calibration-Based Diagnostic for High-Sparsity Post-Pruning Allocation

At very high sparsity, neural network pruning does more than decide which weights remain. It also determines where pruning induced damage is placed across the network, and whether that damage can be recovered by a fixed lightweight repair procedure. We study this problem through the lens of repair conditioned sparsity allocation. We introduce Relative Repairability (RR), a calibration based diagnostic that compares the raw activation distortion caused by layerwise pruning with the residual distortion left after channelwise variance matching repair. RR estimates the fraction of local damage that remains after repair, using only unlabeled calibration data. Across ResNet18, ResNet34, and VGG16 BN on CIFAR10 and CIFAR100, we find that RR is not a universally dominant allocation rule. Instead, it is most useful near an architecture dependent recoverability transition, where standard structural or magnitude based allocation priors begin to lose reliability but post repair recovery has not yet fully collapsed. On CIFAR100 ResNet18, a fine grained sweep shows that RR improves over ERK across the central transition band and surpasses LAMP near the upper part of this band. A projection forced ablation further shows that capped ERK can over protect projection layers, shifting excessive sparsity onto regular convolutions and reducing post repair recovery. These results suggest that high sparsity pruning should allocate not only retained weights, but also repairable damage.

cs.LG↗

How Sparsity Allocation Shapes Label-Free Post-Pruning Recoverability

Unstructured magnitude pruning at high sparsity can reduce neural network accuracy to near-random performance, while labeled retraining may be unavailable in practical deployment settings. Label-free post-pruning repair methods can partially recover collapsed sparse models, but their effectiveness depends on the sparse model left by the upstream pruning allocation. This paper studies how sparsity allocation shapes post-repair recoverability under a fixed activation-statistic repair backend. We compare ERK and LAMP allocations under the same label-free repair protocol across CIFAR-10, CIFAR-100, and Imagenette with ResNet-18, ResNet-34, and ResNet-50 at sparsities from 90% to 95.5%. The results show that allocation choice can substantially change post-repair accuracy at the same global sparsity, and that the preferred allocation varies with architecture, dataset difficulty, and sparsity level. We identify a repair-sensitive transition regime in which BatchNorm recalibration begins to fail, while activation-statistic repair still recovers nontrivial accuracy. Additional validation on ImageNet-100 and DenseNet-121 shows that the location and width of this recoverable regime depend on data scale and connectivity structure. These findings suggest that pruning allocation and post-pruning repair should be studied jointly, since the allocation determines how much activation signal remains available for label-free recovery.

cs.LG↗