arXiv subjects

Shiwen Ni

Publications and source records attributed to Shiwen Ni.

At least 19 recordsLinked to original sources

arXiv2026

Why Generate a Click? Plug-in Decision Heads for Web Agents

Web agents built on vision language models decide every action by generating text such as click(12), although most of that text encodes a choice among a few action types and a few dozen marked elements. We read these choices directly from the hidden states that the generator computes while it reads the goal and the screenshot. Plug-in decision heads on a frozen generator predict the action type, pool the visual tokens inside each numbered box to pick the target, and point at a listed text option or at a WRITE slot that leaves the writing of new text to the generator. A confidence cascade executes the heads' action when they are sure and otherwise lets the generator decode from the KV cache of the same forward pass. With Qwen3-VL 4B and 8B generators trained on 213k web steps, a strict threshold keeps the cascade between 0.3 points below and 1.4 points above the generator on seven offline test sets while the heads decide 12 to 64% of the steps. A looser threshold raises step success on unseen Mind2Web websites by 3.1 and 1.8 points, costs up to 2 points on GUIAct and MiniWoB++, and lifts WebArena task success from 12.5 to 13.8% (4B) and from 13.6 to 14.6% (8B) with less than half of the generated tokens. Training the heads jointly with the generator lowers the generator's own step success by up to 4.9 points; heads on a frozen generator avoid this cost.

cs.CL↗

arXiv2026

SAFE-MR: Evidence Sufficiency Learning for Selective Multimodal Rumor Detection

Multimodal rumor detectors increasingly rely on retrieved evidence, yet relevant evidence is not necessarily sufficient for verification. Missing provenance, duplicated reports, and unresolved contradictions can produce confident predictions without adequate support. We introduce SAFE-MR, a framework that separates claim veracity from evidence sufficiency. The method decomposes image-text posts into verifiable claims, constructs a relation-aware claim-evidence graph, and aggregates evidence using provenance and contextual compatibility. Separate veracity and sufficiency heads support selective prediction, while evidence interventions encourage stability under irrelevant additions and sensitivity to evidence removal. On NewsCLIPpings, VERITE, and XFacta, SAFE-MR achieves macro-F1 scores of 91.2%, 75.8%, and 85.2%, respectively. Against the matched backbone with evidence, its macro-F1 gains are 2.2, 4.9, and 4.8 percentage points. On the diagnostic selection set, SAFE-MR reduces AURC from 0.105 for maximum-probability rejection to 0.075 and lowers error at 80% coverage from 13.8% to 8.5%. Evidence-perturbation and ablation results support the role of sufficiency learning and intervention training in improving selective verification.

cs.CL↗

arXiv2026

LLM-Based Test Generation: Information Sources, Generation Strategies, and Quality Evidence

Large language models are increasingly used to generate test scenarios, executable test suites, assertions, interaction sequences, and fuzzing infrastructure. These artifacts serve different purposes and rely on different sources of information about correct behavior. A test that increases implementation coverage, a test that agrees with a reference program, and a test that detects a requirements violation therefore provide distinct kinds of evidence. We present a structured narrative survey that connects the generation process to the evidence used to justify test quality. Drawing on 95 curated source records through 29 September 2026, we organize the field along four dimensions: testing objectives and artifacts, information available during training and generation, generation and learning mechanisms, and evaluation evidence. We synthesize work on unit and requirements-based testing, test oracles, API and GUI testing, fuzzing, and learned test generators. The resulting framework explains how execution feedback can improve executability while also affecting the independence of an oracle, why specification provenance and training-time reference supervision matter for comparisons, and how downstream code-selection gains differ from test correctness. We use these distinctions to organize benchmarks and derive a research agenda for independent oracle assessment, budget-aware generation, repository-scale evaluation, and test maintenance.

cs.SE↗

arXiv2026

Certified Selective Automation of LLM Agent Evaluation

Evaluating LLM agents still ends with a human reading trajectories, because automatic judges carry no guarantee on how often they are wrong. We ask the operational question: what fraction of agent evaluation can a judge take over, with a certificate that the error rate among auto-decided trajectories stays below a budget alpha? Agent corpora resist the standard answer: many agents attempt the same tasks, so trajectories arrive in correlated clusters, and the i.i.d. certificates of existing selective-judging methods can overstate what is safe: a naive certificate can claim 98% automation while its realized error exceeds the budget in 17.5% of task resamples. We introduce a task-level bootstrap certificate that is valid in every regime we test while matching the naive certificate's coverage; finite-sample cluster-valid alternatives certify nothing at realistic task counts. Under this certificate, a 4B logprob judge trained with SFT and reject-weighted GRPO certifies 0.30-0.59 of evaluation on tool-use and web corpora at alpha=0.1, the only judge, among strongly elicited frontier models, certifying on both headline corpora. Certified coverage is predictable before training from base rate and discrimination alone (leave-one-corpus-out R^2=0.96). Finally, the certificate doubles as a self-training filter: pseudo-labels harvested inside certified regions have contamination bounded by alpha by construction (realized 0.000-0.041 across six harvests), letting a judge enter an unseen domain at in-domain strength with zero target training labels.

cs.CL↗

arXiv2026

Guides That Cause Actions: An Offline Study of Guide-Action Mutual Reinforcement in Multimodal Web Agents

Web agents are usually evaluated in live environments, where environment state and judge models drift between runs, so the same checkpoint rarely reproduces the same score, making controlled studies of training phenomena impractical. We present WebMRE, an offline benchmark of 541 tasks and 5,293 steps derived from successful WebArena trajectories, with fully audited test labels and a deterministic protocol that scores a checkpoint identically on every run without any environment. Each step pairs a human oriented guide sentence with a grounded action, enabling the first study of the mutual reinforcement effect between them in web agents. Averaged over three seeds the effect holds for both models in both decoding orders and grows with scale: jointly decoding a guide lifts element selection over an action only reference by 0.9 and 0.2 points for Qwen3.5-4B and by 1.7 and 2.2 points for Qwen3.5-9B. A mediation analysis shows that the guide is a causal channel rather than commentary: forcing the gold guide as a decoding prefix lifts action accuracy from .422 to .684, another step's guide collapses it to .055, and a paraphrase that renames the target still recovers half of the gain, so the channel carries instruction meaning and not only the label string. The same channel yields an offline reward that only a replayable protocol makes computable, though optimizing it from a strong checkpoint brings no gain yet. Our fine tuned models outperform GPT-5.5, Claude Opus 4.8, and Gemini 3.5 Flash, run zero shot, on every offline metric.

cs.CL↗

arXiv2026

VibeMemBench: Evaluating Memory Systems for Coding Agents on Real Repository Coding Tasks

Coding agents operate on real repository coding tasks, and persistent memory systems promise to reuse experience across tasks. Yet existing evaluations do not show whether those systems improve executable repository work. Repository benchmarks test code changes but do not isolate memory, while memory benchmarks score recall without measuring downstream coding outcomes. We introduce VibeMemBench, a benchmark for evaluating memory systems on 111 coding targets from 90 SWE-rebench V2 repositories and 3,634 history trajectories from the target repositories. The targets follow the SWE benchmark style and cover bug fixes, feature requests, interface changes, and configuration work. An agent edits each target codebase under a declared memory condition. Executable tests decide task resolution. Each target is retained only when injected history experience improves its executable outcome in a reference setting, so every target carries a prior experience whose usefulness is verified by execution in that setting. The frozen verified experience is then transferred to five held-out solvers. Direct injection raises observed task resolution on four of them by 1.1 to 4.5 percentage points while lowering agent steps on all five. Yet when four existing memory systems must construct and retrieve experience from the same history, eleven of twelve solver and system pairings fail to exceed the matched memory-off baseline. VibeMemBench exposes the gap between the useful experience that repository history holds and the experience existing memory systems deliver for repository coding tasks.

cs.SE↗

arXiv2026

Test-Driven Approaches to Software Engineering with Large Language Models: A Survey of Phases, Tasks, and Agent Skills

Tests increasingly participate in the decisions made by large language models and software engineering agents. They specify intended behavior, guide program construction and repair, select candidates, constrain transformations, and provide execution evidence for software analysis. These uses draw on test-driven development, yet differ substantially in test order, oracle availability, editable artifacts, and the role of execution. We present a structured scoping survey organized around the question of what decision a test changes. The review integrates 87 research and supporting records, with method- or protocol-level extraction for 83 records, alongside a separate collection of five practice resources. We distinguish the Red--Green--Refactor cycle from test-conditioned generation, execution-guided refinement, test-mediated analysis, and evaluation-only testing. We then compare code generation, repair, translation, refactoring, clone detection, code search, localization, training-data construction, and formal-specification validation. A dedicated analysis examines how agent workflows and reusable skills encode testing procedures and how their effects are evaluated. Across these tasks, the evidence supports treating test availability, test validity, feedback use, and evaluation independence as separate properties. Test passing alone does not establish behavioral equivalence, effective feedback, or process adherence; aggregate improvements can also conceal different outcomes across models, tasks, and denominators. We synthesize these distinctions into a mechanism taxonomy, a cross-task comparison, and a protocol-sensitive evidence analysis, and identify research directions in oracle validation, causal evaluation, long-horizon maintenance, and reusable test-driven agent capabilities

cs.SE↗

arXiv2026

Do Code Language Models Follow Tests? Paired Interventions on Program Behavior

Visible tests specify concrete program behavior, but an improvement in benchmark accuracy does not establish that a model follows the rule expressed by those tests. We study test utilization through matched prompting controls, paired semantic interventions, and test suites selected by fault detection. Our semantic intervention holds an underspecified description and its example inputs fixed while changing the correct outputs to express one of two valid rules. Evaluation on unseen inputs measures whether both generated programs follow their respective rules. Across five models and three runs of 120 paired instances from 20 specification families, mean switching rates range from 11.1\% to 65.8\%. Qwen3.8-27B has the highest point estimate, followed by Qwen3.6-27B at 60.6\%; their paired difference remains uncertain. Explicit descriptions elicit both rules from these two models on every instance, exposing a gap between implementation capability and adoption of test-specified rules. Correct expected outputs improve MBPP+ accuracy beyond inputs alone for all five models. On 180 tasks with fixed three-test suites, high-detection suites detect 30.3 percentage points more errors in a held-out pool dominated by reference mutants. The corresponding generation differences range from $-0.6$ to $+1.1$ points; all intervals include zero and remain compatible with some benefit. Paired interventions make test-specified rule changes measurable alongside implementation capability and benchmark correctness.

cs.SE↗

arXiv2026

Correct Tests Are Not Enough: Measuring and Training Oracle Conversion in Specification-Based Test Generation

Generating tests from a natural-language specification requires both an input that exposes faulty behavior and a correct expected output. These requirements need not improve together: a model can increase test correctness by choosing easier inputs, or discover useful inputs whose expected outputs it cannot predict. We study this interaction through executable reward decomposition and suite-level oracle-conversion measurement. Our generator jointly emits five input--output tests in one response. During training, audited reference programs provide correctness feedback, while a fixed bank of faulty programs provides two utility signals: potential input kill and effective kill after checking the generated output. An additive GRPO objective preserves both signals without requiring execution at inference time. On an audited TC-Bench split with 506 training and 142 evaluation tasks, three independently trained Qwen3.5-9B runs at step 75 increase full-test correctness from 28.59\% to 42.54\%, input kill from 24.06\% to 25.27\%, and effective full kill from 12.23\% to 14.15\%. Matched 50-step ablations reveal a trade-off: removing kill rewards yields higher correctness and slightly higher full kill, but lowers input kill to 21.60\%. A fixed-input source--oracle crossover on 64 training-pool tasks attributes the principal NoKill-to-FullKill difference to harder input selection rather than worse output prediction on identical inputs. These results identify oracle conversion as a measurable bottleneck and show the benefits and limits of preserving input-utility feedback in joint test generation.

cs.SE↗

arXiv2026

How Output Format Confounds Data Quality and Capability in Instruction Tuning

Instruction-tuning data are judged by quality metrics, and tuned models are judged by benchmarks, but both judgments pass through an output interface: the surface format in which an answer is written. Using gradient signatures across 12 tasks, four semantically equivalent interfaces, three model families, and controlled corruptions, we show that this interface confounds both measurements. Spectral statistics such as effective rank are provably invariant to interface rotation and empirically blind to semantic corruption, while the direction of the update carries the quality signal. The interface-varying residual is not noise: it identifies each unit's own target task perfectly across all three families. Capability itself is stored relative to the training interface: a skill that raises accuracy by more than 40 points under the training format can be nearly invisible under every other, and correcting a single generation budget flips the measured effect of fine-tuning on GSM8K from a gain into a large loss. Pre-registered interventions delimit where this geometry stops short of control. Data quality and model capability are interface-conditioned quantities, and current practice often reports the interface instead of the content.

cs.CL↗

arXiv2026

Joint Training Is Not Enough: Conditioned Cross-Granularity Training for Multimodal Document Understanding

The Mutual Reinforcement Effect (MRE) asks whether a fine, span-level and a coarse, document-level task help each other when one model handles both. We test it in multimodal document understanding on three corpora, two of receipts and one of scanned business forms, comparing single-task, joint and conditioned training, which puts one granularity's gold output in the other's prompt during training only. We build Doc-MRE, an annotation layer pairing gold field extraction (point) with four document-level facets (line), from a three-judge LLM committee under a pre-registration, validated by blind re-annotation. One predicate, fixed in advance: at a shared recipe, a regime reinforces if it beats the matched single-task model on both granularities. Mixed joint training, the arrangement prior MRE work assumes, reinforces on no corpus at the main scale: it is below both single-task models on CORD and trades one granularity for the other on the two others, as single-task tuning does. Conditioned training reinforces on two of the three, CORD (+0.5 point, +4.8 line) and the forms corpus (+7.2 point, +11.0 line), resolvably on the coarse side and directionally on the fine one, and trades on WildReceipt; at that recipe no alternative measurably beats it on either side anywhere. Two byte-identical-prompt controls separate content from format: shuffled conditioning destroys the coarse-side skill but costs the fine side far less, and a neutral-content control reproduces the whole fine-side gain on WildReceipt, which is therefore prompt structure but buys nothing resolvable on the other two. On the forms corpus conditioning buys collapse avoidance: mixed training and the neutral control both assign the majority semantic label to all 50 test documents; only conditioning recovers the gold distribution. Probes find the information decodable under every regime with no resolvable increase under conditioning.

cs.CL↗

arXiv2026

Auditing and Decomposing Feedback-Driven Evolution in LLM Test Generation under the Oracle Problem

Execution feedback is often treated as a self-verifying signal for improving LLM-generated tests. However, when generated inputs are executed on a single accepted program and its outputs are used as ground truth, invalid or underspecified inputs can create spurious fault detections and apparent evolutionary gains. We audit this failure mode in feedback-driven test generation using 142 development tasks, 114 locked external tasks, and 138 held-out tasks, with two code models, three seeds, and fault-cross-fitted real submissions. On external inputs for which three accepted implementations agree, generated outputs match the panel on only 27.79% and 50.12% of cases. A single-reference oracle inflates the measured gain from evolution by 9.46-14.85 percentage points; after auditing, equal-budget independent resampling outperforms mutation-based evolution by 6.01-18.83 points. We further compare a genuine three-round feedback loop with a density-matched placebo. External Real-Placebo differences are +0.13 and -0.50 points, while held-out differences are +1.99 and +0.28 points and do not provide robust evidence of fine-grained feedback benefit. A blinded semantic audit by two software engineering doctoral students classifies 94.41% of panel-disconfirmed inputs as invalid but 3.60% as valid, showing that panel disagreement is informative but not semantic proof. We propose an audit-and-placebo protocol that separates verifier artifacts, interaction scaffolding, and grounded feedback credit in evaluations of self-evolving test generators.

cs.SE↗

arXiv2026

Security Tests as Executable Specifications for LLM Code Generation: Benefits, Trade-offs, and Coverage Limits

Large language models (LLMs) can generate functionally useful code that remains vulnerable, while security-focused interventions may break intended behavior. We investigate security tests as executable specifications both before generation and during iterative repair. We develop SecTDD, a controlled test-feedback scaffold that separates three factors: whether tests are shown upfront, whether failed executions trigger revision, and how failures are selected and represented. The evaluation uses behavior-partitioned visible and hidden tests and byte-identical initial candidates for repair comparisons. Across 2,705 trajectories, 31 task instances, three secure-code benchmarks, 16 CWE categories, and two model families, showing all visible tests upfront increases hidden functional-and-security joint success by 19.3 percentage points on average, but improves only seven of nine benchmark-model conditions and harms two. In shared-candidate comparisons, structured feedback repairs 80 initially unsuccessful candidates with no joint regressions; fixed raw feedback repairs 83 but causes three regressions. Structured and raw feedback are otherwise nearly indistinguishable head-to-head (six wins, six losses, and 453 ties). Candidates that pass all visible tests still fail hidden behavior families under every common regime. These results show that executable feedback can repair secure-code generation, but its benefits depend on the model, task, feedback entry point, and especially test coverage.

cs.SE↗

arXiv2026

HNR-DAC: Hard-Negative Reranking and Distribution-Aligned Classification for Scientific Claim Verification

Scientific claim verification over a cited paper requires predicting the claim--paper relation and identifying the paragraphs that justify that prediction. This setting poses two linked challenges: within-paper distractors often resemble genuine evidence, while a classifier trained on gold evidence must operate on retrieved evidence at inference. We present HNR-DAC, a two-stage framework that trains each stage on the cases it will actually encounter. Hard-Negative Reranking (HNR) quantifies evidence confusability using a base reranker's scores on non-gold paragraphs and contrasts gold evidence against the most confusable candidates. Distribution-Aligned Classification (DAC) trains on the Top-1 paragraph produced by the same frozen HNR used to construct inference inputs, while HNR's Top-3 paragraph identifiers provide the evidence output. On the NLPCC 2026 Task 10 Track 2, the final configuration obtains 97.21% Hit@3, 95.79% Macro-F1, 94.47% Joint@3, and an average score of 95.13%. The corresponding submission ranks third on the official Track 2 leaderboard while achieving the highest overall Macro-F1 of 93.05%, alongside 70.16% Joint@3 and an average score of 81.61%.

cs.CL↗

arXiv2026

Probing the Difficulty Perception Mechanism of Large Language Models

Large language models (LLMs) are increasingly deployed on complex reasoning tasks, yet little is known about their ability to internally evaluate problem difficulty, which is an essential capability for adaptive reasoning and efficient resource allocation. In this work, we investigate whether LLMs implicitly encode problem difficulty in their internal representations. Using a linear probe on the final-token representations of LLMs, we demonstrate that the difficulty level of math problems can be linearly modeled. We further locate the specific attention heads of the final Transformer layer: these attention heads have opposite activation patterns for simple and difficult problems, thus achieving perception of difficulty. Our ablation experiments prove the accuracy of the location. Crucially, our experiments provide practical support for using LLMs as automatic difficulty annotators, potentially substantially reducing reliance on costly human labeling in benchmark construction and curriculum learning. We also uncover that there is a significant difference in entropy and difficulty perception at the token level. Our study reveals that difficulty perception in LLMs is not only present but also structurally organized, offering new theoretical insights and practical directions for future research. Our code is available at https://github.com/Aegis1863/Difficulty-Perception-of-LLMs.

cs.CL↗

arXiv2026

MAG: A Web-Agent Benchmark and Harness for Multimodal Action and Guide Generation

Digital Adoption Platforms (DAPs) are embedded overlays widely used on web systems to guide users through operations inside a page, helping them get started with unfamiliar interfaces quickly. Completing a real task, however, rarely means clicking a few buttons on a single page: it takes a sequence of actions that unfolds across changing page states. Prior studies have also treated automated web agent actions and guide text generation as two separate problems, and most of them feed models textual page representations such as the DOM or accessibility trees rather than the rendered screens that humans actually operate on. In this work we introduce MAG, the first benchmark that unifies task execution and guide writing into a single Multimodal Action and Guide task, with two grounding schemes over screenshots: Set-of-Mark element selection and raw pixel coordinates. We further build a complete harness for this compound task, covering annotation with LLM assistance and human verification, training, evaluation in live environments, and joint metrics for actions and guides. With this harness we evaluate frontier API models and open multimodal models, and report detailed analyses. Finally, we design a GRPO training method augmented with expert trajectories, which nearly doubles the success rate of a supervised 9B agent (from 6.9% to 13.2%) and improves guide quality at the same time. Even the strongest model completes fewer than 40% of the tasks, leaving ample room for future research.

cs.AI↗

arXiv2026

A Learning-Rate-Gated Failure of GRPO in a Small Language and Vision-Language Model Web Agent: A Controlled Null and Its Mechanism

Reinforcement learning with verifiable rewards, and Group Relative Policy Optimization (GRPO) in particular, is now run routinely on a supervised checkpoint in the hope of producing a stronger agent. We ask whether it adds skill to a small language and vision-language model web agent at the 4B to 8B scale, or whether it mostly reshapes behavior the supervised model already has. Across a control grid of 18 runs that varies learning rate, KL weight, seed, initialization, and clipping, no configuration credibly improves the success rate of a strong supervised baseline on tasks the agent has largely mastered. On the text track, moderate to high learning rates make it credibly worse. The null holds under paired testing, 25 evaluation seeds, 6 training seeds, changes to the recipe, both text and Set-of-Marks screenshot observations, and scaling the backbone to 8B; the credible harm is a text-track finding and is only nominal under Set-of-Marks. To show that the null reflects the setting and not a broken pipeline, we run the identical harness, reward, and recipe on tasks whose reward is reachable by sampling, and there the success rate rises by 22 points with a paired interval that excludes zero. GRPO therefore helps only when there is headroom to climb, meaning the sampled policy already succeeds more often than the greedy one. We then explain the failure. A middle learning rate degrades the agent and a high one collapses it, and the two regimes form a double dissociation: grafting localizes the degrade regime to the attention and MLP blocks, while the collapse regime cannot be traced to any single group, and the embedding change that dominates the weight movement is causally inert. At 4B, effective rank in the late layers tracks capability in both directions; at 8B the two come apart. This coupling is specific to the smaller model, so we report it as scale-dependent.

cs.AI↗

arXiv2026

SrDetection: A Self-Referential Framework for Data Leakage Detection in Code Large Language Models

Evaluating code large language models (Code LLMs) requires reliable detection of data leakage, where benchmark performance is artificially inflated by exposure to benchmark data during pre-training. Existing approaches either assume access to proprietary training corpora, rely on brittle heuristics such as timestamp filtering, or use external reference sets with manually tuned, non-generalizable thresholds. To address these limitations, we introduce \textbf{SrDetection}, a unified \textbf{s}elf-\textbf{r}eferential leakage detection framework for both gray-box (access to model logits) and black-box (access to model outputs) settings. SrDetection generates semantically equivalent variants of a benchmark sample and detects leakage by contrasting the model's behavior on the original versus its variants, flagging cases where the original is disproportionately easier for the model. We further design a controlled leakage detection testbed and evaluate SrDetection in this environment. Across different models and training stages, SrDetection improves average F1 by 21.52 points in the gray-box setting and 14.46 points in the black-box setting over strong baselines, demonstrating robust, threshold-independent leakage detection. Finally, a gray-box study of 15 widely used Code LLMs on four popular benchmarks reveals benchmark-specific leakage patterns beyond prior overlap-based analyses\footnote{\footnotesize Source code and data are available at https://github.com/SMinL/SrDetectionCode

cs.CL↗