arXiv subjects

Ziheng Peng

Publications and source records attributed to Ziheng Peng.

8 recordsLinked to original sources

arXiv2026

Do LLMs Know Their Vulnerable Scenarios?

Safety-aligned large language models are trained to refuse harmful requests, yet embedding the same requests in particular scenarios can bypass their safeguards. Existing red-teaming methods empirically identify effective scenarios through observed attack outcomes, but why particular scenarios weaken refusal remains mechanistically unclear. Meanwhile, mechanistic interpretability studies have characterized both refusal directions and jailbreak-associated features, without explaining the relationship between the two representations. In this work, we show that scenario-wrapped prompts activate internal scenario directions whose causal steering consistently reduces refusal scores. Building on this finding, we propose \textsc{Concept2Scenario}, a concept-based attribution framework for vulnerable scenario discovery. It instantiates a broad concept space with a sparse autoencoder, attributes refusal suppression to individual concepts, translates the identified concepts into interpretable natural-language scenarios, and identifies synergistic scenario combinations through interaction attribution. Across three open-source models, two safety benchmarks, and six black-box jailbreak methods, the discovered scenarios serve as reusable priors that improve average attack success rates by up to $18.2$ percentage points. They also transfer to GPT-5, Claude-Haiku-4.5, and Gemini-3-Flash, suggesting that some scenario-level refusal vulnerabilities are shared across model families. Moreover, the identified combinations outperform their individual constituents and enable iterative attacks to succeed in fewer turns.

cs.AI↗

arXiv2026

SkillEval: Decomposing Agent Skill Quality into Interpretable Signals

Agent skills provide reusable procedural knowledge that helps agents solve specialized tasks. As their use expands, evaluating skill quality becomes increasingly important. Existing evaluations often measure skill quality by testing whether a skill improves performance on specific downstream tasks. However, a reusable skill may apply to multiple task scenarios. Downstream evaluation mainly reflects the compatibility between a skill and the evaluated task, provides only a partial view of skill quality, and does not identify which aspect of the skill should be improved. We find that general properties of the \texttt{SKILL.md} document play an important role in skill quality. To evaluate these properties, we propose \textbf{SkillEval}, an interpretable framework for document-level skill evaluation. SkillEval evaluates each property using a fixed and inspectable scoring direction, producing interpretable scores. It further measures and reduces the influence of unrelated document features, such as length and formatting, so that each score captures its intended semantic property more specifically. Specifically, SkillEval learns an interpretable direction for each quality property from controlled positive--negative skill pairs in the hidden representation space of the model, and scores a new skill by projecting its representation onto these fixed directions. We use SkillEval to evaluate skills in controlled quality tests and show that SkillEval reliably distinguishes skills of different quality. In addition, SkillEval scores closely reflect downstream task performance, providing an early indication of whether a skill is likely to help an agent complete a task. We further explore SkillEval for diagnosing weaknesses in skill documents and guiding targeted revisions. The revised skills improve the targeted properties and achieve higher pass rates on downstream tasks.

cs.AI↗

arXiv2026

Self-Improving Large Language Models via Progressive Experience Evolution

Large language models (LLMs) capable of self-improvement require not only effective policy optimization, but also a principled mechanism for transforming transient interaction experience into persistent model capabilities. Existing self-improvement paradigms remain fragmented: test-time methods can explicitly extract experience but cannot internalize it into model parameters, whereas training-time optimization methods can update model parameters but lack an explicit mechanism for accumulating transferable experience. Bridging these two paradigms requires a critical intermediate stage that remains underexplored, namely \emph{experience distillation}. To address this gap, we propose \textbf{SPEE} (\textbf{S}elf-\textbf{P}rogressive \textbf{E}xperience \textbf{E}volution), a unified post-training framework that sequentially performs explicit experience evolution followed by implicit policy optimization. During explicit experience evolution, SPEE reflects on trajectories collected from multiple interactions to extract, verify, and progressively evolve transferable experience, which is subsequently internalized into the policy through privilege-guided On-Policy Self-Distillation (OPSD). During implicit policy optimization, reward-driven reinforcement learning leverages these internalized priors to explore novel solution strategies. In the experience evolution stage, a continuously evolving global experience pool consolidates knowledge from both successful and failed trajectories, filters out low-utility experience, and mitigates post-hoc rationalization induced by individual trajectories. Experiments on five mathematical reasoning benchmarks demonstrate that SPEE consistently outperforms both test-time and training-time self-evolution baselines across three model scales. The source code is available at https://github.com/rrrsj/SPEE.

cs.CL↗

arXiv2026

To Call or Not to Call: Diagnosing Intrinsic Over-Calling Bias in LLM Agents

LLM agents exhibit a consistent tendency to over-call, invoking tools even in situations where none is needed. On the When2Call benchmark, six models from three families show high call accuracy but much lower no-call accuracy, leaving overall accuracy in the 55%-70% range. We trace this to an Intrinsic Bias Hypothesis (IBH): the call/no-call decision mapping carries an activation-independent call offset, so the model favors call even at activation parity. Using Sparse Autoencoders (SAEs), we recover behavior-aligned feature bases for the call/no_call decision, reduce them to a signed activation margin, and estimate the offset directly. Across all six models, the model is decision-neutral only when no_call activation outweighs call activation, consistent with IBH. We then causally test IBH with Adaptive Margin-Calibrated Steering (AMCS), a closed-form counter-bias shift along SAE decoder directions. Cancelling the diagnosed offset mitigates over-calling and improves overall accuracy with a negligible drop in call accuracy. Our work recasts over-calling from an empirical phenomenon into a mechanistic object amenable to causal correction. Code is available at https://github.com/SKURA502/agent-sae/.

cs.LG↗

arXiv2026

Target Concept Tuning Improves Extreme Weather Forecasting

Deep learning models for meteorological forecasting often fail in rare but high-impact events such as typhoons, where relevant data is scarce. Existing fine-tuning methods typically face a trade-off between overlooking these extreme events and overfitting them at the expense of overall performance. We propose TaCT, an interpretable concept-gated fine-tuning framework that solves the aforementioned issue by selective model improvement: models are adapted specifically for failure cases while preserving performance in common scenarios. To this end, TaCT automatically discovers failure-related internal concepts using Sparse Autoencoders and counterfactual analysis, and updates parameters only when the corresponding concepts are activated, rather than applying uniform adaptation. Experiments show consistent improvements in typhoon forecasting across different regions without degrading other meteorological variables. The identified concepts correspond to physically meaningful circulation patterns, revealing model biases and supporting trustworthy adaptation in scientific forecasting tasks. The code is available at https://anonymous.4open.science/r/Concept-Gated-Fine-tune-62AC.

cs.LG↗

arXiv2026

ProtoTS: Learning Hierarchical Prototypes for Explainable Time Series Forecasting

While deep learning has achieved impressive performance in time series forecasting, it becomes increasingly crucial to understand its decision-making process for building trust in high-stakes scenarios. Existing interpretable models often provide only local and partial explanations, lacking the capability to reveal how heterogeneous and interacting input variables jointly shape the overall temporal patterns in the forecast curve. We propose ProtoTS, a novel interpretable forecasting framework that achieves both high accuracy and transparent decision-making through modeling prototypical temporal patterns. ProtoTS computes instance-prototype similarity based on a denoised representation that preserves abundant heterogeneous information. The prototypes are organized hierarchically to capture global temporal patterns with coarse prototypes while capturing finer-grained local variations with detailed prototypes, enabling expert steering and multi-level interpretability. Experiments on multiple realistic benchmarks, including a newly released LOF dataset, show that ProtoTS not only exceeds existing methods in forecast accuracy but also delivers expert-steerable interpretations for better model understanding and decision support.

cs.LG↗

arXiv2025

DiffSeg30k: A Multi-Turn Diffusion Editing Benchmark for Localized AIGC Detection

Diffusion-based editing enables realistic modification of local image regions, making AI-generated content harder to detect. Existing AIGC detection benchmarks focus on classifying entire images, overlooking the localization of diffusion-based edits. We introduce DiffSeg30k, a publicly available dataset of 30k diffusion-edited images with pixel-level annotations, designed to support fine-grained detection. DiffSeg30k features: 1) In-the-wild images--we collect images or image prompts from COCO to reflect real-world content diversity; 2) Diverse diffusion models--local edits using eight SOTA diffusion models; 3) Multi-turn editing--each image undergoes up to three sequential edits to mimic real-world sequential editing; and 4) Realistic editing scenarios--a vision-language model (VLM)-based pipeline automatically identifies meaningful regions and generates context-aware prompts covering additions, removals, and attribute changes. DiffSeg30k shifts AIGC detection from binary classification to semantic segmentation, enabling simultaneous localization of edits and identification of the editing models. We benchmark three baseline segmentation approaches, revealing significant challenges in semantic segmentation tasks, particularly concerning robustness to image distortions. Experiments also reveal that segmentation models, despite being trained for pixel-level localization, emerge as highly reliable whole-image classifiers of diffusion edits, outperforming established forgery classifiers while showing great potential in cross-generator generalization. We believe DiffSeg30k will advance research in fine-grained localization of AI-generated content by demonstrating the promise and limitations of segmentation-based methods. DiffSeg30k is released at: https://huggingface.co/datasets/Chaos2629/Diffseg30k

cs.CV↗

arXiv2024

CORE: Mitigating Catastrophic Forgetting in Continual Learning through Cognitive Replay

This paper introduces a novel perspective to significantly mitigate catastrophic forgetting in continuous learning (CL), which emphasizes models' capacity to preserve existing knowledge and assimilate new information. Current replay-based methods treat every task and data sample equally and thus can not fully exploit the potential of the replay buffer. In response, we propose COgnitive REplay (CORE), which draws inspiration from human cognitive review processes. CORE includes two key strategies: Adaptive Quantity Allocation and Quality-Focused Data Selection. The former adaptively modulates the replay buffer allocation for each task based on its forgetting rate, while the latter guarantees the inclusion of representative data that best encapsulates the characteristics of each task within the buffer. Our approach achieves an average accuracy of 37.95% on split-CIFAR10, surpassing the best baseline method by 6.52%. Additionally, it significantly enhances the accuracy of the poorest-performing task by 6.30% compared to the top baseline. Code is available at https://github.com/sterzhang/CORE.

cs.LG↗