Searcharxiv⌕ Search

arXiv subjects

Yiwei Chen

Publications and source records attributed to Yiwei Chen.

At least 19 recordsLinked to original sources

ACED-Bench: Evaluating How LLM Agents Acquire and Act on Causal Evidence

Large language models are increasingly used as scientific assistants, yet we still lack reliable ways to evaluate whether they can investigate a system rather than merely answer a prompt. A useful evaluation must let agents choose experiments, interpret finite evidence, and make causal decisions, while preserving objective ground truth. We introduce ACED (Active Causal Evidence and Decision Benchmark), a neuro-symbolic benchmark that separates the readable scientific surface from the formal causal world: natural-language scenarios describe the system, while a hidden causal model defines the data, the effect of every intervention, and the oracle answers. In ACED-Repair, agents diagnose a failing production system: within 15 experiments they must decide which controls to test, read confounded and filtered data, and recommend a fix, which we re-simulate on a hidden true objective. In ACED-Query, agents answer causal questions and make constrained causal decisions in worlds backed by categorical probabilistic graphical models. Acceptance on ACED-Repair ranges from 43% to 0% across 12 LLMs, while strong agents approach ceiling on ACED-Query's qualification questions. Tracing every experiment shows where investigations stop: most open-weight models test a true lever on its own in fewer than half of the worlds, the strongest models test the right levers but often miss the working dose or combination, and agents that do run a working fix almost always recommend it. No model reliably repairs systems with two independent causes or systems that need a subgroup-specific policy. These results suggest that reliable AI co-scientists need explicit support for choosing experiments, keeping track of evidence, and composing interventions. Code and data are available at https://github.com/Rose-STL-Lab/ACED-Bench.

cs.AI↗

CVE-Conditioned Exploit Generation with LLMs: A Data-Centric Study of Benchmarking and Fine-Tuning

We study the task of CVE-conditioned exploit generation, where a model drafts proof-of-concept (PoC) exploits given software vulnerability context. We adopt a data-centric approach, constructing a high-quality dataset via multi-stage preprocessing and introducing a scalable evaluation framework with LLM-as-judge and fine-grained rubrics. Under this unified setup, we benchmark 17 large language models across 8 evaluation criteria, providing systematic insights into their zero-shot capabilities. We further show that a compact 8B open-weight model, when fine-tuned on curated data, achieves over 42.5% improvement in exploit quality and rivals some proprietary models when combined with simple test-time rejection strategies. Our results highlight the importance of data quality, structured supervision, and evaluation design for reliable exploit generation, suggesting that these factors can be as critical as model scale in adapting LLMs to cybersecurity tasks.

cs.CR↗

Who Built This Model? Tracing LLM Lineage via Spectral Fingerprints in Weight Space

Open-weight large language models (LLMs) are increasingly developed through complex, multi-stage pipelines, leading to intricate lineage relationships that reflect model origin, ownership, and evolution. Understanding these relationships is important for model provenance, governance, and supply-chain integrity. In this work, we investigate the notion of LLM "biometrics" (analogous to human biometrics) to ask whether LLMs exhibit intrinsic fingerprints in weight space alone, without access to input data, that reveal their origin and lineage. We formulate this as a lineage discrimination problem, distinguishing among independent-origin, same-series, and shared-base models. To characterize these relationships, we propose a unified geometric fingerprinting framework that analyzes weight matrices from two complementary perspectives: (i) spectral energy, captured by singular value distributions to encode global magnitude patterns, and (ii) subspace alignment, quantified via subspace deviations to capture directional geometry. Our analysis uncovers a clear hierarchy of structural similarity in weight space: spectral energy reliably distinguishes independently trained models and different model families, while subspace alignment enables fine-grained discrimination among closely related models, including variations in dataset scale and post-training procedures. Extensive experiments on over 110 diverse open-weight LLM pairs demonstrate that weight-space geometry provides a robust and interpretable signal for model lineage, enabling coarse-grained regime separation and fine-grained discrimination within shared-base models.

cs.AI↗

Can Native Safety Travel? Reusing Safety Pretraining Across Models via On-Policy Distillation

Safety pretraining has emerged as a promising paradigm for embedding safety deeply into a model during pretraining, yielding robustness to subsequent fine-tuning that conventional safety post-training often lacks. We refer to this safety acquired during model pretraining as native safety. Yet native safety is expensive to acquire, confined to the model in which it was pretrained, and often accompanied by lower general utility. In this work, we rethink the practical value of safety-pretrained models and introduce, for the first time, the native safety reuse problem: Can native safety be transferred from a specialized, potentially smaller, safety-pretrained "teacher" to standard-pretrained "students," enabling them to inherit its robustness while preserving their general utility? We show that native safety can "travel" through on-policy distillation. A standard-pretrained student generates responses to safety-pretraining prompts, while a small safety-pretrained teacher supervises the student's token-level predictions along its own generated trajectories; cross-tokenizer alignment further enables transfer across model families. The resulting students exhibit substantially greater safety robustness than their safety-post-trained counterparts while retaining higher general utility than the teacher: subsequent fine-tuning raises the average attack success rate (ASR) of the 1.7B student by only 5.00%, compared with 11.58% for safety post-training, and its average utility of 48.20% far exceeds the aligned teacher's 27.42%. Unlike conventional large-to-small knowledge distillation, our framework enables selective, small-to-large safety transfer: it reuses the robust safety expertise of a smaller pretrained teacher without inheriting its utility limitations. These findings establish native safety as a reusable model property rather than one permanently tied to the originally pretrained model.

cs.LG↗

Train Locally, Learn Together: Distributed Multi-Agent Training via ADMM

Multi-agent systems (MAS) can outperform individual language-model agents by combining specialized roles, diverse reasoning processes, and collaborative decision-making. However, collaborative MAS training is typically centralized, requiring a common trainer to aggregate role-local trajectories and learning signals and maintain model and optimization states. This motivates a distributed multi-agent learning paradigm in which agents improve system-level capabilities through role-local learning and decentralized communication, paving the way toward scalable, resilient, and self-organizing MAS. To this end, we introduce, to the best of our knowledge, the first optimization-grounded framework for distributed MAS training based on consensus alternating direction method of multipliers (ADMM). We formulate collaborative MAS learning as a consensus-constrained finite-sum optimization problem, naturally lending itself to distributed optimization via ADMM. The proposed approach alternates role-local learning with inter-agent consensus using distributed communication protocols such as ring all-reduce, enabling joint optimization while keeping each agent's data and training state local. We evaluate our framework on distributed safety alignment and general reasoning tasks. Under compromised-agent settings, ADMM can reduce HarmBench attack success rate (ASR) by up to 26.4%, while achieving reasoning gains comparable to centralized training. These results show that distributed MAS training retains the benefits of centralized joint optimization without requiring a centralized trainer.

cs.AI↗

Visual Abstention in Unified Multimodal Models

Unified multimodal models (UMMs) integrate understanding and generation, yet their generative behavior is rarely governed by what they understand about the task. We formalize visual abstention: when a requested visual transformation is impossible under the task's rules, the model should recognize that no valid solution exists, state this, and decline to generate. We introduce Draw-or-Decline (DoD), a benchmark of 1,050 feasible-infeasible request pairs across 7 task categories that jointly measures editing success and the refusal of infeasible requests. Evaluating 8 UMMs, we find that editing ability and abstention are distinct capabilities: even the strongest editor, at 68.4% editing accuracy, refuses only 0.4% of infeasible requests under ordinary instructions. Their reasoning shows why: the models rarely notice the conflict, and instead plan the edit as if the request were possible, often describing objects that are not in the image, or quietly change the request into one they can complete. Explicitly prompting these UMMs to report infeasibility increases textual refusals but reduces editing accuracy. We propose VisTA (Visual Transformation and Abstention), a training method that pairs feasible and infeasible examples so that a model judges feasibility before deciding whether to generate. We train VisTA-BAGEL to perform feasible edits and decline infeasible requests. Without any reminder, it refuses 93.0% of infeasible requests, up from 0.4% for the strongest editor, while falsely refusing only 0.8% of feasible ones. Unlike a reminder, this does not cost editing accuracy: VisTA-BAGEL completes 74.3% of feasible edits, more than any of the 8 evaluated UMMs.

cs.CL↗

E-WAVE: Event-based Continuous Optical Flow via Warping-Aligned Visual Encoding

Temporally dense optical flow is essential for dynamic perception in immersive VR/AR systems, where rapid head, hand, and object motion must be continuously captured and tracked. Existing frame-based optical flow estimation methods are constrained by the tradeoff between temporal resolution and computational cost; while event cameras, with their high temporal resolution and energy efficiency, serve as a natural solution to the dilemma. However, event-based approaches commonly rely on correlation volumes to capture pairwise voxel correspondences, which incur substantial memory and computation overhead. We present E-WAVE, a correlation-free framework for high-temporal-resolution (HTR) optical flow estimation from event streams. Instead of constructing all-pairs correlation volumes, E-WAVE employs global attention mechanism to model long-range feature dependencies and performs trajectory guided feature warping using Bézier curve. Through iterative updates, it predicts trajectories that allow for querying at arbitrary timestamps without repeated inference. Experiments on MultiFlow and DSEC-Flow demonstrate a 25% lower trajectory error and comparable endpoint flow estimation accuracy relative to state-of-the art baselines. Additional evaluations on self-captured data using a head-mounted prototype validate that E-WAVE remains robust under challenging real-world conditions.

cs.CV↗

Forgetting to Forget: Attention Sink as A Gateway for Backdooring LLM Unlearning

Large language model (LLM) unlearning is a key approach for removing undesired data, knowledge, or behaviors from pretrained models while retaining their general utility. Yet, with the rise of open-weight LLMs, we ask: can the unlearning process itself be backdoored, appearing successful under normal conditions yet reverting to pre-unlearned behavior when a hidden trigger is activated? Drawing inspiration from classical backdoor attacks that embed triggers into training data to enforce specific behaviors, we investigate backdooring unlearning, a setting in which models forget as intended in the clean setting but recover forgotten knowledge when the trigger appears. We show that designing such attacks presents unique challenges, hinging on where triggers are placed and how backdoor training is reinforced. We uncover a strong link between the backdoor efficacy and the attention sink phenomenon (i.e., shallow input tokens consistently attract disproportionate attention). Our analysis reveals that these attention sinks serve as gateways for backdooring unlearning: placing triggers at sink positions and aligning their attention values markedly enhances backdoor persistence. Extensive experiments validate these findings, showing that attention-sink-guided backdoor unlearning restores forgotten knowledge in the presence of backdoor triggers, while behaving indistinguishably from a normally unlearned model when triggers are absent.

cs.LG↗

Emergence of Topological Electron Crystals in Bilayer Graphene--Mott Insulator Heterostructures

The interplay between strong electron correlation and band topology offers a playground for discovering exotic quantum phases. Here, we predict the emergence of topological electron crystals in a charge-transfer bilayer graphene-Mott insulator heterostructure. In this system, interlayer charge transfer induces a charge-neutral electron-hole bilayer with strong mass asymmetry. While the extreme dilute limit favors a classical triangular dipolar Wigner crystal, we show that increasing the carrier density triggers a critical competition between the Coulomb interaction and the underlying topological band structure of bilayer graphene. This interplay destabilizes the triangular dipolar Wigner crystal and instead stabilizes intrinsic quantum electron crystals with spontaneously formed honeycomb and kagome geometries. Crucially, these new phases can host distinct topological responses including the quantum anomalous and quantum spin Hall effects, which inherit the nonlocal quantum geometry of the bilayer graphene wave functions. Our results establish this artificial heterostructure as a highly tunable platform for mimicking two-dimensional topological solids in a single device.

cond-mat.mes-hall↗

Possible inverse magnetic melting effect in vdW-like Kondo lattice CeSn$_{0.75}$Sb$_2$

Given the intimate connection between magnetic orders and the interplay among multiple degrees of freedom in heavy-fermion systems, controlling and understanding the associated inverse melting effect is crucial for unveiling novel condensed-matter states and their potential applications. Here, we report the growth of single crystalline quasi-two-dimensional van-der-Waals-like (vdW-like) Kondo lattice CeSn$_{0.75}$Sb$_2$, and its physical properties by a combination of transport / magnetic / thermodynamic measurements. We find that it hosts a fragile antiferromagnetic (AFM) order and a cluster glass (CG) ground state, both of which are highly sensitive to external fields. Upon cooling under low in-plane magnetic fields, the AFM phase evolves into a polarized paramagnetic phase, either directly or indirectly through the intermediate CG phase. This process constitutes a possible inverse magnetic melting effect that restores the broken translational / rotational symmetries. Our work provides a rare paradigm of inverse magnetic melting effect in vdW-like heavy-fermion materials, and enriches the physics in conventional Kondo-lattice models.

cond-mat.str-el↗

Chosen-Plaintext Attacks of Double Random Phase Encryption with Nonlinear Optical Media

This paper studies an inverse problem in nonlinear optical encryption. We examine chosen-plaintext attacks (CPA) on a nonlinear optical encryption strategy that integrates double random phase encryption (DRPE) into a nonlinear optical propagation model to enhance the security of the combined system. We first demonstrate that the system's phase information can be decoded from carefully designed differential CPA data. We then demonstrate that the strength of the optical device's nonlinearity can also be recovered from CPA data, indicating that including this parameter as an additional security key does not enhance protection against CPA attacks, although numerical simulations show that strong nonlinearity still poses significant challenges for CPA attacks. Finally, we provide a stability analysis to demonstrate that small errors in decoded security keys result in only small errors in the decrypted text, even though the encryption process is nonlinear.

physics.optics↗

Safety Mirage: How Spurious Correlations Undermine VLM Safety Fine-Tuning and Can Be Mitigated by Machine Unlearning

Recent vision language models (VLMs) have made remarkable strides in generative modeling with multimodal inputs, particularly text and images. However, their susceptibility to generating harmful content when exposed to unsafe queries raises critical safety concerns. While current alignment strategies primarily rely on supervised safety fine-tuning with curated datasets, we identify a fundamental limitation we call the ''safety mirage'', where supervised fine-tuning inadvertently reinforces spurious correlations between superficial textual patterns and safety responses, rather than fostering deep, intrinsic mitigation of harm. We show that these spurious correlations leave fine-tuned VLMs vulnerable even to a simple one-word modification-based attack, where substituting a single word in text queries with a spurious correlation-inducing alternative can effectively bypass safeguards. Additionally, these correlations contribute to the over-prudence, causing fine-tuned VLMs to refuse benign queries unnecessarily. To address these issues, we show machine unlearning (MU) as a powerful alternative to supervised safety fine-tuning, as it avoids biased feature-label mappings and directly removes harmful knowledge from VLMs while preserving their general capabilities. Extensive evaluations across safety benchmarks show that under MU-based alignment reduces the attack success rate by up to 60.27% and cuts unnecessary rejections by over 84.20%. WARNING: There exist AI generations that may be offensive in nature.

cs.AI↗

Unlearning Isn't Invisible: Detecting Unlearning Traces in LLMs from Model Outputs

Machine unlearning (MU) for large language models (LLMs), commonly referred to as LLM unlearning, seeks to remove specific undesirable data or knowledge from a trained model, while maintaining its performance on standard tasks. While unlearning plays a vital role in protecting data privacy, enforcing copyright, and mitigating sociotechnical harms in LLMs, we identify a new vulnerability post-unlearning: unlearning trace detection. We discover that unlearning leaves behind persistent "fingerprints" in LLMs, detectable traces in both model behavior and internal representations. These traces can be identified from output responses, even when prompted with forget-irrelevant inputs. Specifically, even a simple supervised classifier can determine whether a model has undergone unlearning, using only its prediction logits or even its textual outputs. Further analysis shows that these traces are embedded in intermediate activations and propagate nonlinearly to the final layer, forming low-dimensional, learnable manifolds in activation space. Through extensive experiments, we demonstrate that unlearning traces can be detected with over 90% accuracy even under forget-irrelevant inputs, and that larger LLMs exhibit stronger detectability. These findings reveal that unlearning leaves measurable signatures, introducing a new risk of reverse-engineering forgotten information when a model is identified as unlearned, given an input query.

cs.LG↗

Leak@$k$: Unlearning Does Not Make LLMs Forget Under Probabilistic Decoding

Unlearning in large language models (LLMs) is critical for regulatory compliance and for building ethical generative AI systems that avoid producing private, toxic, illegal, or copyrighted content. Despite rapid progress, in this work, we show that \textit{almost all} existing unlearning methods fail to achieve true forgetting in practice. Specifically, while evaluations of these `unlearned' models under deterministic (greedy) decoding often suggest successful knowledge removal using standard benchmarks, we show that sensitive information reliably resurfaces when models are sampled with standard probabilistic decoding. To rigorously capture this vulnerability, we introduce \texttt{leak@$k$}, a new meta-evaluation metric that quantifies the likelihood of forgotten knowledge reappearing when generating $k$ samples from the model under realistic decoding strategies. Using three widely adopted benchmarks, TOFU, MUSE, and WMDP, we conduct the first large-scale, systematic study of unlearning reliability using \texttt{leak@$k$} metric. Our findings demonstrate that knowledge leakage persists across methods and tasks, underscoring that current state-of-the-art (SOTA) unlearning techniques provide only limited forgetting. We propose an algorithm, termed Robust Unlearning under LEak@$k$ metric (\texttt{RULE}) to address this concern. We demonstrate that \texttt{RULE} provides an unlearned model for TOFU benchmark with no information leakage for a large number of generation samples. On the MUSE benchmark, \texttt{RULE} outperforms SOTA unlearning methods under the \texttt{leak@$k$} metric across most sampling budgets $k$. Codes are available at https://github.com/OptimAI-Lab/Leak-k.

cs.LG↗

Interlaced R2D2 DNN Series for Scalable Non-Cartesian MRI with Sensitivity Self-calibration

We introduce interlaced R2D2 (iR2D2), a DNN series paradigm for scalable image reconstruction from accelerated non-Cartesian k-space acquisitions in MRI with sensitivity map self-calibration. While unrolled DNN architectures provide robust image formation, embedding non-uniform fast Fourier transform operators within the backpropagation graph becomes impractical to train at large scale, e.g., in 2D MRI with a large number of coils, or for higher-dimensional imaging. To address this scalability challenge, we leverage the R2D2 paradigm as a learned version of the Matching Pursuit algorithm that was recently introduced in radio astronomy for fast large-scale Fourier imaging. R2D2's reconstruction is formed as a series of residual images iteratively estimated as outputs of DNN modules taking the previous iteration's data residual as input. Specific to MRI, precomputed sensitivity maps derived from undersampled data can yield an inaccurate measurement operator, which may adversely affect the performance of iterative algorithms such as R2D2. Thus, we extend the R2D2 framework to iR2D2 by introducing a bespoke interlaced architecture that alternates between two R2D2 DNN series to jointly self-calibrate sensitivity maps and form the MR image. We further enhance iR2D2 to operate as an adaptive solver governed by an error-controlled update condition that enforces a sufficient residual energy descent, a dynamic capability fundamentally incompatible with the predefined forward passes of unrolled architectures. Extensive experiments in simulation and on real data, targeting undersampled radial k-space sampling, demonstrate that iR2D2 significantly improves upon R2D2 and outperforms state-of-the-art benchmarks, delivering scalable, high-fidelity imaging with corrected sensitivity profiles.

eess.IV↗

Less Is More: Fast and Accurate Reasoning with Cross-Head Unified Sparse Attention

Large reasoning models achieve strong performance through test-time scaling, but this incurs substantial computational overhead due to long decoding from short prompts. While sparse attention can reduce latency and memory usage, existing methods often degrade reasoning accuracy because selection errors accumulate over long generation horizons, or require costly retraining. We introduce LessIsMore, a training-free sparse attention mechanism for long-horizon reasoning. Our key insight is that token importance in reasoning is global and stable: critical tokens are largely shared across attention heads and remain stable over decoding steps. Guided by this structure, LessIsMore enforces cross-head unified token selection and preserves recent context via a stable recency window, yielding a globally consistent token set that can be reused across layers. Across multiple model families and challenging reasoning benchmarks, LessIsMore matches or improves accuracy while attending to substantially fewer tokens. With kernel-level optimizations, LessIsMore achieves up to $1.6\times$ end-to-end decoding speedup and up to $1.72\times$ faster sparse attention computation, with additional long-context results demonstrating the generality of our approach. Code is available at \href{https://github.com/DerrickYLJ/LessIsMore}{https://github.com/DerrickYLJ/LessIsMore}.

cs.CL↗

Zephyrus: An Agentic Framework for Weather Science

Foundation models for weather science are pre-trained on vast amounts of structured numerical data and outperform traditional weather forecasting systems. However, these models lack language-based reasoning capabilities, limiting their utility in interactive scientific workflows. Large language models (LLMs) excel at understanding and generating text but cannot reason about high-dimensional meteorological datasets. We bridge this gap by building the first agentic framework for weather science. Our framework includes a Python code-based environment for agents (ZephyrusWorld) to interact with weather data, featuring tools including a WeatherBench 2 dataset indexer, geolocator for geocoding from natural language, weather forecasting, climate simulation capabilities, and a climatology module for querying precomputed climatological statistics (e.g., means, extremes, and quantiles) across multiple timescales. We design Zephyrus, a multi-turn LLM-based weather agent that iteratively analyzes weather datasets, observes results, and refines its approach through conversational feedback loops. We accompany the agent with a new benchmark, ZephyrusBench, with a scalable data generation pipeline that constructs diverse question-answer pairs across weather-related tasks, from basic lookups to advanced forecasting, extreme event detection, and counterfactual reasoning. Experiments on this benchmark demonstrate the strong performance of Zephyrus agents over text-only baselines, outperforming them by up to 44 percentage points in correctness. However, the hard tasks are still difficult even with frontier LLMs, highlighting the challenging nature of our benchmark and suggesting room for future development. Our codebase and benchmark are available at https://github.com/Rose-STL-Lab/Zephyrus.

cs.AI↗

ORCA: Object Recognition and Comprehension for Archiving Marine Species

Marine visual understanding is essential for monitoring and protecting marine ecosystems, enabling automatic and scalable biological surveys. However, progress is hindered by limited training data and the lack of a systematic task formulation that aligns domain-specific marine challenges with well-defined computer vision tasks, thereby limiting effective model application. To address this gap, we present ORCA, a multi-modal benchmark for marine research comprising 14,647 images from 478 species, with 42,217 bounding box annotations and 22,321 expert-verified instance captions. The dataset provides fine-grained visual and textual annotations that capture morphology-oriented attributes across diverse marine species. To catalyze methodological advances, we evaluate 18 state-of-the-art models on three tasks: object detection (closed-set and open-vocabulary), instance captioning, and visual grounding. Results highlight key challenges, including species diversity, morphological overlap, and specialized domain demands, underscoring the difficulty of marine understanding. ORCA thus establishes a comprehensive benchmark to advance research in marine domain. Project Page: http://orca.hkustvgd.com/.

cs.CV↗