arXiv subjects

Yifeng He

Publications and source records attributed to Yifeng He.

At least 19 recordsLinked to original sources

arXiv2026

BenchShield: Formal Model-Backed Instrumentation for Reward Integrity in LLM-Agent Evaluation Infrastructure

LM-agent benchmarks increasingly function as interactive evaluation infrastructure. Agents observe state, call tools, modify workspaces, submit artifacts, and receive rewards from outcome procedures. This interactivity makes evaluations vulnerable to reward hacking: an agent improves its measured score by exploiting the reward-relevant trajectory instead of solving the intended task. Existing defenses rely largely on task-specific patches, prompt instructions, or post-hoc detectors. They do not provide reusable evidence that a concrete run remained within its intended evaluation boundary. This paper presents BenchShield, a model-backed instrumentation layer for reward integrity in LLM-agent evaluation. BenchShield grounds detection in a finite lifecycle model of an evaluation's reward-relevant events. Within the benchmark infrastructure, two complementary analyses operate over this model. A static, phase-aware taint analysis exposes reward-hacking paths before a run. Its runtime counterpart uses infrastructure-side evidence to attribute concrete agent use and emit evidence-backed claims. We construct BenchShield Trajectories, a human-labeled corpus of 456 adjudicated trajectories from more than 31,000 public agent runs across three benchmarks. Compared with an agentic hackability scanner baseline on the same tasks and model, BenchShield improves full-chain recall from 23-94% to 77-100%, same-vector coverage from 16-56% to 43-78%, and reduces per-task cost by up to 65%. Its runtime analysis achieves 96% accuracy in detecting reward hacking from infrastructure-side evidence.

cs.CR↗

arXiv2026

Invariant Pretraining for Robust Code Representations

Encoder-based code representation models remain widely deployed for discriminative tasks such as clone detection and code classification, where their small size and low inference cost are decisive. Their robustness, however, is fragile: under invariant programs, semantically equivalent code written in different syntactic forms, learned representations degrade substantially even though program behavior is unchanged. We present an empirical study of this robustness gap across four encoder baselines, two downstream tasks, and four datasets, together with a minimal code-only continued pretraining recipe that closes much of it. Our method, invariant pretraining (InvPT), applies semantic-preserving transformations to the corpus and combines masked language modeling with multi-positive supervised contrastive learning that treats all augmentations of the same source function as positives, mixing self-contrast pairs (same code, different masks) with invariant-contrast pairs (transformed code) for positives of varying difficulty. Unlike prior contrastive code encoders, InvPT does not require paired natural-language data. Across our evaluation, InvPT improves robustness on transformed test sets by up to 11 percentage points on clone detection and 19 on code classification while matching or improving standard accuracy, and our ablations isolate multi-positive invariant contrast as the main source of the gains. Our aim is not a new objective but a careful measurement of where encoder robustness breaks and how far a simple, code-only recipe can recover it.

cs.LG↗

arXiv2026

The Greatness of Science Cannot Be Planned: Agentic Auto-Research is Fuzz Testing

Agentic auto-research is emerging, but most systems treat scientific discovery as goal-oriented optimization against a final benchmark. This paradigm rewards a sparse final verdict and ignores the exploration that precedes it. When agents optimize only the final score, they overfit to the test conditions and sample blindly rather than search. Within a declared research problem, a research agent and a greybox fuzzer for software analysis face the same sparse feedback. A fuzzer rarely finds a bug directly, but coverage makes partial progress observable on every execution. Fuzzers use that dense signal to mutate inputs and allocate effort, rather than merely rank completed runs. Auto-research needs the same two capabilities. First, each experiment must expose a cheap, dense signal of epistemic progress before final scientific validation is available. Second, that signal must determine the next intervention so the agent searches rather than repeatedly samples. Because the progress signal provides guidance rather than a final verdict, final validation must still evaluate claims using evidence protected from adaptive reuse. We propose controlled tests to determine whether candidate signals predict validated progress, whether feedback-directed search yields more validated discoveries per unit cost than repeated sampling, and whether protected validation reduces false discoveries. In a simulated physics environment, an AI research agent that tracks its intermediate epistemic progress discovers a hidden physical law. Optimization-driven baselines fail because they repeatedly sample and overfit to their existing data instead of probing unfamiliar regimes. Feedback architecture, not generation capacity, is the central bottleneck in auto-research.

cs.AI↗

arXiv2026

LLM-Based Invariant Testing for Software Functional Bugs

Manually writing unit tests to uncover functional bugs in software libraries is not only time-consuming but also requires a deep understanding of the intended semantics of the APIs. Heuristic-based test generation methods suffer from low usability because they cannot reason about program semantics or interpret source code and documentation as humans do. Traditional fuzzing techniques like OSS-Fuzz often rely on crashes to detect bugs, but functional bugs do not always cause crashes. To overcome these limitations, we present LISA, a novel LLM-based invariant testing framework for software functional bugs. LISA iteratively generates API sequences and program invariants guided by API n-gram feedback, achieving higher bug-detection rates and competitive code coverage compared with both fuzzing and prior LLM-based test generation approaches, and reporting each finding as a high-confidence bug candidate for developer confirmation.

cs.SE↗

arXiv2026

Is Progressive Disclosure All You Need for Long-Context Agents?

Long-document question answering usually forces a choice between loading the whole document into the context window and bolting on a separate retriever. Agentic AI suggests a broader option, giving the agent the document path and letting it decide how and what to read. Agent Skills, a standard for packaging expertise into folders an agent loads on demand, supply a ready mechanism: progressive disclosure, which exposes only what a query needs, from a short description down to the specific passages. Practitioners rapidly adopted this pattern for book-length understanding tasks, but the evidence to support such choices has been anecdotal. We run the first controlled study of the pattern, comparing raw-document navigation and several designs of Agent Skills packs against a classical hybrid retriever across three agent harnesses and three model families on InfiniteBench. On a single book, the gain depends on the harness, running large when the agent navigates the raw document poorly but near zero when a strong agent harness already divides and retrieves on its own. When scaling up to tasks that span many books, raw-document navigation collapses while one-level progressive disclosure degrades more slowly and pulls ahead. A second, deeper routing level never helps and sometimes breaks accuracy outright, so one level is enough. Progressive disclosure buys context, not intelligence: it is redundant while a strong agent can locate the right passages itself, and decisive once the corpus grows too large to navigate by reading.

cs.AI↗

arXiv2026

LightCrafter: PBR-Conditioned Video Diffusion Refinement for Controllable and Consistent Relighting

Video relighting requires balancing long-form temporal consistency with a physically grounded understanding of light transport, which depends on accurate estimation of intrinsic scene properties such as materials, geometry, and illumination. Existing methods follow two paradigms: (1) reconstruct a video's photometric properties via inverse rendering and relight them to a target illumination via forward rendering, using physically-based rendering (PBR) or a neural renderer; these suffer from noisy reconstructions and struggle with hard-to-model effects such as global illumination. (2) Frame the task as generative video-to-video translation conditioned on relighting targets (a target environment map or text); this limits relighting control and temporal stability, since diffusion models struggle to translate long-form videos, and is constrained by the availability of input/relit training pairs. We propose LightCrafter, a hybrid pipeline that reformulates video relighting as video translation of a proxy video: rather than translating the input video directly to the target, we translate a PBR rendering of the input under the target illumination to the final target. This bakes illumination targets into the PBR proxy, removing the need to teach the diffusion model illumination concepts like environment maps, and enables more intricate lighting control while naturally providing long-form temporal consistency. We show PBR renders alone already outperform some prior art but struggle with effects like global illumination; to capture these, we leverage photometric priors in video generation models by post-training CogVideoX on synthetic video pairs and real-world unpaired videos. We outperform prior state-of-the-art on existing real-world relighting benchmarks and contribute a synthetic benchmark for further analysis. We will release our dataset, benchmark, metrics, and code.

cs.CV↗

arXiv2026

Code Generation by Differential Test Time Scaling

Test-time scaling has emerged as a promising approach for improving code generation by exploring large solution spaces at inference time. However, existing methods often rely on public test cases that are unavailable in practice, or require extensive LLM inference for candidate selection, leading to significant token consumption and time overhead. We present DiffCodeGen, a novel test-time scaling method for code generation based on coverage-guided differential analysis. DiffCodeGen generates diverse code candidates using various sampling and prompting strategies, then applies coverage-guided fuzzing to synthesize inputs without requiring any existing tests or large language models. By executing all candidates on these inputs, DiffCodeGen captures their dynamic behavior and clusters candidates based on behavioral similarity. DiffCodeGen selects the medoid of the largest cluster as the final output. Unlike prior test-time scaling methods that invoke additional LLM inference for candidate selection, DiffCodeGen performs selection without any extra model calls, incurring little to no additional token consumption. DiffCodeGen is fully asynchronous, naturally suited to the current trend of agentic coding, and is thus efficient and highly scalable. We evaluate DiffCodeGen across 4 large language models, demonstrating consistent improvements over baselines. Compared to state-of-the-art test-time scaling methods, DiffCodeGen achieves competitive or superior performance while using only a fraction of time and tokens. DiffCodeGen is model-agnostic and can be combined with reasoning models to further boost performance.

cs.SE↗

arXiv2026

ContractBench: Can LLM Agents Preserve Observation Contracts?

Tool-augmented LLM agents call APIs whose intermediate outputs, such as presigned URLs, session tokens, and OAuth state parameters, are observation contracts: artifacts whose later use is constrained by the external system that produced them. We show that observation contract compliance (preserving the temporal validity and byte-level integrity) is an emergent, regression-prone capability: it is neither guaranteed by general tool-use ability nor consistently improved by larger or newer models. To measure this, we introduce ContractBench, a benchmark of 33 dual-axis tasks that probe two orthogonal failure modes no existing benchmark evaluates: validity failures (using an artifact after expiry) and integrity failures (corrupting an artifact's bytes through the observation-to-action pipeline). Our evaluation is deterministic and programmatic, with a virtual clock controlling time and SHA-256 hashes verifying byte integrity. We assign each outcome a failure label drawn from real-world API specifications. We evaluate 38 models and report four findings: (i) no evaluated model clears 80%, with Claude-Opus-4.6 leading at 77.8%, revealing that current frontier models still fail to comply with observation contracts; (ii) a sharp within-family capability cliff in Qwen 3.5 between 4B (0%) and 9B (56.6%), smoothing to 70.7% at 397B-A17B: what emerges across the cliff is mid-trajectory restraint, not tool-call competence; (iii) non-monotonic scaling across the GPT-5 family: agentic post-training can erode compliance through sycophancy-driven regression; (iv) our failure taxonomy works as an actionable in-context reward signal, yielding +7.1 pp on 42 paired GPT-5.1 failures.

cs.SE↗

arXiv2026

Content Fuzzing for Escaping Information Cocoons on Digital Social Media

Information cocoons on social media limit users' exposure to posts with diverse viewpoints. Modern platforms use stance detection as an important signal in recommendation and ranking pipelines, which can route posts primarily to like-minded audiences and reduce cross-cutting exposure. This restricts the reach of dissenting opinions and hinders constructive discourse. We take the creator's perspective and investigate how content can be revised to reach beyond existing affinity clusters. We present ContentFuzz, a confidence-guided fuzzing framework that rewrites posts while preserving their human-interpreted intent and induces different machine-inferred stance labels. ContentFuzz aims to route posts beyond their original cocoons. Our method guides a large language model (LLM) to generate meaning-preserving rewrites using confidence feedback from stance detection models. Evaluated on four representative stance detection models across three datasets in two languages, ContentFuzz effectively changes machine-classified stance labels, while maintaining semantic integrity with respect to the original content.

cs.CL↗

arXiv2026

SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks

Agent Skills are structured packages of procedural knowledge that augment large language model (LLM) agents at inference time. Despite rapid adoption, there is no standard way to measure whether they actually help. We present SkillsBench, a benchmark whose current inventory contains 87 tasks across 8 domains paired with curated Skills and deterministic verifiers. Our latest aggregate evaluation runs the 87-task benchmark under matched no-Skills and curated-Skills conditions for 18 model-harness configurations. Curated Skills raise the average pass rate from 33.9% to 50.5% (+16.6 percentage points; 25.5% normalized gain), with configuration-level gains ranging from +4.1 to +25.7 pp. Focused Skills with at most three modules outperform larger or exhaustive bundles, and smaller models with Skills can match larger models without them. SkillsBench establishes paired evaluation as the foundation for rigorous measurement of Skill efficacy on agentic, expertise-heavy work.

cs.AI↗

arXiv2025

Formation of external particle jets on a spherical particle bed subjected to strong explosive loading

We report the mechanism for the formation of external particle jets on a spherical particle bed subjected to strong explosive loading, revealing a critical dependence on particle size. Under strong explosive loading, the formation of external particle jets is primarily driven by a drag-coupled mechanism. We conducted Eulerian-Lagrangian simulations, with up to $2048^3$ effective cells and $1.8$ million tracked parcels on an adaptive mesh, for both small- and large-particle cases. Pronounced jets are observed only with small particles, alongside accelerated bed thickening. By defining characteristic inner and outer radii, the particle bed thickness evolution is quantified, showing an initial linear growth followed by a nonlinear deceleration. Particle dynamics analysis indicates that drag force dominates particle motion and jet formation during the nonlinear stage. The initial angular non-uniformity of the particle bed induces a non-uniform gas radial velocity. Through drag coupling, this flow asymmetry generates a radial velocity difference in small particles, thereby promoting pronounced jet formation, whereas large particles resist this drag-induced effect. The greater drag-induced deceleration on smaller particles leads to an increased velocity difference across the particle bed, explaining the accelerated thickening. A characteristic radius model that integrates the Gurney model for the linear stage with a drag-dominated deceleration model for the nonlinear stage is established and shows good agreement with numerical results across different particle sizes.

physics.flu-dyn↗

arXiv2025

SpecTokenizer: A Lightweight Streaming Codec in the Compressed Spectrum Domain

Neural Audio Codecs (NACs) have gained growing attention in recent years as technologies for audio compression and audio representation in speech language models. While mainstream NACs typically require G-level computation and M-level parameters, the performance of lightweight and streaming NACs remains underexplored. This paper proposes SpecTokenizer, a lightweight streaming codec that operates in the compressed spectral domain. Composed solely of alternating CNN and RNN layers, SpecTokenizer achieves greater efficiency and better representational capability through multi-scale modeling in the compressed spectrum domain. At 4 kbps, the proposed SpecTokenizer achieves comparable or superior performance compared to the codec with state-of-the-art lightweight architecture while requiring only 20% of the computation and 10% of the parameters. Furthermore, it significantly outperforms the codec when using similar computational and storage resources.

eess.AS↗

arXiv2025

Evaluating Program Semantics Reasoning with Type Inference in System F

Large Language Models (LLMs) are increasingly integrated into the software engineering ecosystem. Their test-time compute (TTC) reasoning capabilities show significant potential for understanding program logic and semantics beyond mere token recognition. However, current benchmarks for code reasoning lack a formal, program-centric deductive framework to ensure sound evaluation, and are incapable of assessing whether models genuinely reason about program semantics or merely exploit superficial associations between natural language and code tokens. To bridge this gap, we introduce TF-Bench, a benchmark designed to evaluate LLM reasoning based on type inference in System F, a task we refer to as program semantics reasoning. By employing verified transformations to remove semantically irrelevant natural language, we construct TF-Bench_pure, a purely semantics-driven variant of TF-Bench. Our analysis reveals substantial limitations in state-of-the-art LLMs, with the best-performing LLM (Claude-3.7-sonnet) achieving only 55.85% accuracy on TF-Bench_pure. Additionally, we propose two novel metrics to assess robustness and the effectiveness of test-time reasoning, underscoring critical limitations in current LLM capabilities and highlighting essential directions for future research.

cs.CL↗

arXiv2025

Deep Learning-Enhanced Visual Monitoring in Hazardous Underwater Environments with a Swarm of Micro-Robots

Long-term monitoring and exploration of extreme environments, such as underwater storage facilities, is costly, labor-intensive, and hazardous. Automating this process with low-cost, collaborative robots can greatly improve efficiency. These robots capture images from different positions, which must be processed simultaneously to create a spatio-temporal model of the facility. In this paper, we propose a novel approach that integrates data simulation, a multi-modal deep learning network for coordinate prediction, and image reassembly to address the challenges posed by environmental disturbances causing drift and rotation in the robots' positions and orientations. Our approach enhances the precision of alignment in noisy environments by integrating visual information from snapshots, global positional context from masks, and noisy coordinates. We validate our method through extensive experiments using synthetic data that simulate real-world robotic operations in underwater settings. The results demonstrate very high coordinate prediction accuracy and plausible image assembly, indicating the real-world applicability of our approach. The assembled images provide clear and coherent views of the underwater environment for effective monitoring and inspection, showcasing the potential for broader use in extreme settings, further contributing to improved safety, efficiency, and cost reduction in hazardous field monitoring. Code is available on https://github.com/ChrisChen1023/Micro-Robot-Swarm.

cs.RO↗

arXiv2025

Spontaneous Donor Defects and Voltage-Assisted Hole Doping in Beta-Gallium Oxides under Multiple Epitaxy Conditions

Beta-phase gallium oxide (beta-Ga2O3) is prone to the spontaneous formation of donor defects but poses a formidable challenge in achieving high-quality p-type doping, mainly due to its exceptionally low valence band maximum (VBM). In this study, we utilize first-principles computations to investigate the origin of spontaneous donor defects in beta-Ga2O3 grown by three typical techniques: molecular beam epitaxy (MBE), metal organic chemical vapor deposition (MOCVD), and halide vapor phase epitaxy (HVPE). Our findings elucidate that the primary donor defects vary with the growth techniques, specifically Gai3+ for MBE, Hi+ and CGa+ for MOCVD, and (2VGa+Gai+2VO)+ and ClO+ for HVPE under unintentionally doped conditions. Employing a theoretically proposed voltage-assisted doping method, we computationally demonstrate that the dominant spontaneous donors can be significantly reduced accompanied by a noticeable increase in acceptors, leading to a stepwise reduction of Fermi level to 0.52, 0.88, and 2.10 eV above VBM for the MOCVD, HVPE, and MBE methods, and a hole concentration of 8.5*10^17, 8.7*10^11, and 2.7*10^-9 cm-3, respectively, at room temperature without the use of external dopants. By introducing Mg doping, we further reduce the Fermi level for both the MBE and HVPE experiments.

cond-mat.mtrl-sci↗

arXiv2024

SteerDiff: Steering towards Safe Text-to-Image Diffusion Models

Text-to-image (T2I) diffusion models have drawn attention for their ability to generate high-quality images with precise text alignment. However, these models can also be misused to produce inappropriate content. Existing safety measures, which typically rely on text classifiers or ControlNet-like approaches, are often insufficient. Traditional text classifiers rely on large-scale labeled datasets and can be easily bypassed by rephrasing. As diffusion models continue to scale, fine-tuning these safeguards becomes increasingly challenging and lacks flexibility. Recent red-teaming attack researches further underscore the need for a new paradigm to prevent the generation of inappropriate content. In this paper, we introduce SteerDiff, a lightweight adaptor module designed to act as an intermediary between user input and the diffusion model, ensuring that generated images adhere to ethical and safety standards with little to no impact on usability. SteerDiff identifies and manipulates inappropriate concepts within the text embedding space to guide the model away from harmful outputs. We conduct extensive experiments across various concept unlearning tasks to evaluate the effectiveness of our approach. Furthermore, we benchmark SteerDiff against multiple red-teaming strategies to assess its robustness. Finally, we explore the potential of SteerDiff for concept forgetting tasks, demonstrating its versatility in text-conditioned image generation.

cs.CV↗

arXiv2024

RepoGenReflex: Enhancing Repository-Level Code Completion with Verbal Reinforcement and Retrieval-Augmented Generation

In real-world software engineering tasks, solving a problem often requires understanding and modifying multiple functions, classes, and files across a large codebase. Therefore, on the repository level, it is crucial to extract the relevant information to achieve accurate code completion effectively. Existing code completion tools have achieved some success, but they struggle to optimize the retrieval and generation process dynamically. In this paper, we propose RepoGenReflex, a generic, dynamic, effective framework to address this challenge. By leveraging the Retrieval-Augmented Generation (RAG) enhanced with Verbal Reinforcement Learning (VRL), it can dynamically choose the optimal results for repository-level code completion. RepoGenReflex uses Reflector to give directional feedback to the next loop. RepoGenReflex chooses the optimal results stored in the Experience cache based on the RAG-VRL loop. To validate the framework's generalization ability, we propose a new benchmark RepoGenEval, which consists of the latest, high-quality real-world repositories in line completion scenarios. Our experiments demonstrate that RepoGenReflex achieves significant improvements after optimizing the Reflector component, resulting in enhanced accuracy and relevance of code completions. Additionally, RepoGenReflex consistently demonstrates superior performance and effectiveness across standard code completion tasks, highlighting the robustness and adaptability of our framework.

cs.SE↗

arXiv2024

Error Correction by Paying Attention to Both Acoustic and Confidence References for Automatic Speech Recognition

Accurately finding the wrong words in the automatic speech recognition (ASR) hypothesis and recovering them well-founded is the goal of speech error correction. In this paper, we propose a non-autoregressive speech error correction method. A Confidence Module measures the uncertainty of each word of the N-best ASR hypotheses as the reference to find the wrong word position. Besides, the acoustic feature from the ASR encoder is also used to provide the correct pronunciation references. N-best candidates from ASR are aligned using the edit path, to confirm each other and recover some missing character errors. Furthermore, the cross-attention mechanism fuses the information between error correction references and the ASR hypothesis. The experimental results show that both the acoustic and confidence references help with error correction. The proposed system reduces the error rate by 21% compared with the ASR model.

cs.CL↗