arXiv subjects

Feijiang Han

Publications and source records attributed to Feijiang Han.

12 recordsLinked to original sources

arXiv2026

Meta$^2$: Precomputing Reusable Knowledge for Online Harness Evolution in Long-Horizon Agents

Harness optimization improves long-horizon agents by refining the prompts, tools, memory, and hooks around a frozen model. Offline methods can learn from completed trajectories, but deploy one shared harness that may overfit training tasks and cannot react to evidence revealed during a new run. Online evolution can adapt before, during, or after one run through an evolver, but may have to reason about every intervention from scratch or rely on verifier feedback that is usually unavailable at test time. We introduce Meta$^2$, which moves offline optimization one level up: rather than optimizing a shared harness, it learns reusable knowledge for harness evolution and applies it within each task online. Offline, Meta$^2$ learns a textual Meta-Skill $M$ by contrasting training rollouts of the same task agent under three configurations: without an evolver, with an evolver guided by the current version of $M$, and with an evolver given privileged verifier information available only in training. Contrasts in trajectories and outcomes are aggregated across tasks to refine $M$ into reusable guidance on when and how to evolve, when to leave the harness unchanged, and which interventions to avoid. Online, an evolver combines the frozen $M$ with emerging trajectory evidence to decide whether to KEEP the harness unchanged or EVOLVE it through a task-local edit. Each edit applies to the rest of the same run without resetting its environment or progress. Thus, evolution guidance transfers across tasks while harness edits remain task-local, and held-out verifier feedback is never available to the evolver. On two long-horizon benchmarks, Terminal-Bench 2.1 and DeepSWE 1.1, Meta$^2$ achieves the highest performance among evaluated methods with GPT-5.6-Terra, reaching 81.00% and 44.44%, respectively. Ablations support the value of contrastive evolution guidance and keeping online harness edits task-local.

cs.AI↗

arXiv2026

Non-Gaussianity Reveals Interpretable Directions in Language Models

Finding human-interpretable directions in language-model activations is important for understanding and controlling model behavior. These directions are usually discovered through indirect learning objectives such as sparse reconstruction, but it remains unclear what existing statistical structure distinguishes interpretable directions and whether that structure can be used directly for feature discovery. We find that non-Gaussianity provides such a signature: projection distributions of natural-text residual-stream activations along sparse autoencoder (SAE) decoder directions are substantially more non-Gaussian than those along random directions, with most contexts forming a concentrated background and recurring patterns often occupying one tail. This raises a natural question: if interpretable directions tend to have non-Gaussian projection distributions, can we discover them by searching for non-Gaussian directions directly in activation space, without training another sparse dictionary? To test this hypothesis, we introduce ICA Lens, a practical workflow for efficient discovery, reliable annotation, and faithful evaluation of non-Gaussian directions in LLM residual-stream activations. Across GPT-2 Small, Gemma 2 2B, and Qwen 3.5 9B Base, we find that: (1) the recovered non-Gaussian directions capture coherent lexical, syntactic, and semantic patterns under both human and LLM-based evaluation; (2) non-Gaussianity is positively associated with both human-label confidence and automated interpretability; and (3) the resulting coordinates are competitive with SAEs across interpretability, sparse reconstruction, concept prediction, and targeted probe perturbation, while fitting in minutes per layer and using 7-43 times less memory. These results show that non-Gaussianity is a direct unsupervised signal for discovering interpretable and useful structure in activations.

cs.LG↗

arXiv2026

RealControlBench: A Benchmark for Interpretability-Informed Fine-Grained Control of LLMs and Agents

Existing steering benchmarks are built around simplified, isolated generation tasks and evaluate a single behavioral attribute under controlled settings that do not reflect the diversity and complexity of practical deployment. As a result, performance on these benchmarks provides only a limited picture of how steering methods perform in the real world. We therefore introduce RealControlBench, a comprehensive benchmark for evaluating fine-grained steering of LLMs and LLM agents across diverse behavioral targets and interaction settings. RealControlBench extends steering evaluation beyond isolated generations to complex, execution-grounded, multi-turn agent interactions, covering five complementary dimensions: instruction-following control, agentic tool-use decision control, reasoning-performance steering, reasoning-length control, and trajectory-conditioned safety awareness. Rather than measuring steering success alone, the benchmark jointly evaluates target controllability and capability preservation, including instruction following, reasoning performance, and downstream agent utility and safety. It further examines whether steering effects remain specific and robust across prompt variations, domains, and model architectures. Our evaluation shows that existing steering methods remain unreliable across realistic control objectives: they often fail to consistently induce the intended behavior and can substantially degrade the model's underlying task performance. These results suggest that strong steering effects in controlled settings do not necessarily translate into reliable control in more realistic evaluations. By providing a unified, model-agnostic steering evaluation framework, RealControlBench aims to support the development of steering methods that are not only effective but also precise, robust, and useful in realistic AI deployments.

cs.CL↗

arXiv2026

Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Models

Mechanistic Interpretability (MI) has emerged as a vital approach to demystify the opaque decision-making of Large Language Models (LLMs). However, existing reviews primarily treat MI as an observational science, summarizing analytical insights while lacking a systematic framework for actionable intervention. To bridge this gap, we present a practical survey structured around the pipeline: "Locate, Steer, and Improve." We formally categorize Localizing (diagnosis) and Steering (intervention) methods based on specific Interpretable Objects to establish a rigorous intervention protocol. Furthermore, we demonstrate how this framework enables tangible improvements in Alignment, Capability, and Efficiency, effectively operationalizing MI as an actionable methodology for model optimization. The curated paper list of this work is available at https://github.com/rattlesnakey/Awesome-Actionable-MI-Survey.

cs.CL↗

arXiv2026

ThinknCheck: Grounded Claim Verification with Compact, Reasoning-Driven, and Interpretable Models

We present ThinknCheck, a 1B-parameter verifier for grounded claim verification that first produces a short, structured rationale and then a binary verdict. We construct LLMAggreFact-Think, a 24.1k reasoning-augmented training set derived from LLMAggreFact, and fine-tune a 4-bit Gemma3 model to follow this format. On LLMAggreFact, ThinknCheck attains 78.1 balanced accuracy (BAcc), surpassing MiniCheck-7B (77.4) with 7x fewer parameters; removing the reasoning step reduces BAcc to 57.5. On SciFact, ThinknCheck reaches 64.7 BAcc, a +14.7 absolute gain over MiniCheck-7B. By contrast, zero-shot chain-of-thought on the base Gemma3-1B harms accuracy relative to direct answers, and preference optimization with a simple format+accuracy reward underperforms supervised reasoning. To probe the latter, we introduce GSMClaims and a domain-specialized variant, ThinknCheck-Science, which improves across benchmarks, including 61.0\% accuracy on GSMClaims. Overall, explicit, supervised reasoning enables compact verifiers that are competitive while remaining resource-efficient and interpretable.

cs.AI↗

arXiv2026

DeepAFL: Deep Analytic Federated Learning

Federated Learning (FL) is a popular distributed learning paradigm to break down data silo. Traditional FL approaches largely rely on gradient-based updates, facing significant issues about heterogeneity, scalability, convergence, and overhead, etc. Recently, some analytic-learning-based work has attempted to handle these issues by eliminating gradient-based updates via analytical (i.e., closed-form) solutions. Despite achieving superior invariance to data heterogeneity, these approaches are fundamentally limited by their single-layer linear model with a frozen pre-trained backbone. As a result, they can only achieve suboptimal performance due to their lack of representation learning capabilities. In this paper, to enable representable analytic models while preserving the ideal invariance to data heterogeneity for FL, we propose our Deep Analytic Federated Learning approach, named DeepAFL. Drawing inspiration from the great success of ResNet in gradient-based learning, we design gradient-free residual blocks in our DeepAFL with analytical solutions. We introduce an efficient layer-wise protocol for training our deep analytic models layer by layer in FL through least squares. Both theoretical analyses and empirical evaluations validate our DeepAFL's superior performance with its dual advantages in heterogeneity invariance and representation learning, outperforming state-of-the-art baselines by up to 5.68%-8.42% across three benchmark datasets.

cs.LG↗

arXiv2026

Can LLMs Handle WebShell Detection? Overcoming Detection Challenges with Behavioral Function-Aware Framework

WebShell attacks - where adversaries implant malicious scripts on web servers - remain a persistent threat. Prior machine-learning and deep-learning detectors typically depend on task-specific supervision and can be brittle under data scarcity, rapid concept drift, and out-of-distribution (OOD) deployment. Large language models (LLMs) have recently shown strong code understanding capabilities, but their reliability for WebShell detection remains unclear. We address this gap by (i) systematically evaluating seven LLMs (including GPT-4, LLaMA-3.1-70B, and Qwen-2.5 variants) against representative sequence- and graph-based baselines on 26.59K PHP scripts, and (ii) proposing Behavioral Function-Aware Detection (BFAD), a behavior-centric framework that adapts LLM inference to WebShell-specific execution patterns. BFAD anchors analysis on security-sensitive PHP functions via a Critical Function Filter, constructs compact LLM inputs with Context-Aware Code Extraction, and selects in-context demonstrations using Weighted Behavioral Function Profiling, which ranks examples by a behavior-weighted, function-level similarity. Empirically, we observe a consistent precision-recall asymmetry: larger LLMs often achieve high precision but miss attacks (lower recall), while smaller models exhibit the opposite tendency; moreover, off-the-shelf LLM prompting underperforms established detectors. BFAD substantially improves all evaluated LLMs, boosting F1 by 13.82% on average; notably, GPT-4, LLaMA-3.1-70B, and Qwen-2.5-Coder-14B exceed prior SOTA benchmarks, while Qwen-2.5-Coder-3B becomes competitive with traditional methods. Overall, our results clarify when LLMs succeed or fail on WebShell detection, provide a practical recipe, and highlight future directions for making LLM-based detection more reliable.

cs.CR↗

arXiv2026

ZeroTuning: Unlocking the Initial Token's Power to Enhance Large Language Models Without Training

Token-level attention tuning, a class of training-free methods including Post-hoc Attention Steering (PASTA) and Attention Calibration (ACT), has emerged as a promising approach for improving frozen LLMs via interpretable interventions. However, these methods rely on auxiliary heuristics to identify important task-specific tokens, which can introduce bias and limit applicability when token importance is ambiguous or when optimized kernels make attention maps inaccessible. We propose a simpler alternative: intervening only on the initial token (e.g., BOS in LLaMA). We theoretically show that adding lightweight biases to this token's attention logits systematically shifts and reshapes downstream attention patterns - an effect amplified by its natural role as an attention sink. Empirically, we find that this tuning can improve LLM performance and better elicit pretrained knowledge, with stronger effects in early layers and distinct scaling preferences across attention heads. Building on these findings, we introduce ZeroTuning, a training-free method that improves LLM performance by applying head-specific attention adjustments to the initial token, requiring no parameter updates. We present two variants: a supervised mode that calibrates on validation examples, and an unsupervised mode that directly minimizes output entropy. ZeroTuning requires no KV-cache or decoding changes and is kernel-agnostic (works with SDPA and FlashAttention). It requires only four lines of modification to the standard LlamaAttention code, achieves gains across 15 datasets, and outperforms prior, more complex methods. For example, on Llama-3.1-8B, it yields relative improvements of 19.9% on classification, 4.5% on question answering, and 2.1% on dialogue. ZeroTuning also works out of the box with quantized inference and maintains its improvements as context length increases.

cs.CL↗

arXiv2025

Beyond Detection: A Comprehensive Benchmark and Study on Representation Learning for Fine-Grained Webshell Family Classification

Malicious WebShells pose a significant and evolving threat by compromising critical digital infrastructures and endangering public services in sectors such as healthcare and finance. While the research community has made significant progress in WebShell detection (i.e., distinguishing malicious samples from benign ones), we argue that it is time to transition from passive detection to in-depth analysis and proactive defense. One promising direction is the automation of WebShell family classification, which involves identifying the specific malware lineage in order to understand an adversary's tactics and enable a precise, rapid response. This crucial task, however, remains a largely unexplored area that currently relies on slow, manual expert analysis. To address this gap, we present the first systematic study to automate WebShell family classification. Our method begins with extracting dynamic function call traces to capture inherent behaviors that are resistant to common encryption and obfuscation. To enhance the scale and diversity of our dataset for a more stable evaluation, we augment these real-world traces with new variants synthesized by Large Language Models. These augmented traces are then abstracted into sequences, graphs, and trees, providing a foundation to benchmark a comprehensive suite of representation methods. Our evaluation spans classic sequence-based embeddings (CBOW, GloVe), transformers (BERT, SimCSE), and a range of structure-aware algorithms, including Graph Kernels, Graph Edit Distance, Graph2Vec, and various Graph Neural Networks. Through extensive experiments on four real-world, family-annotated datasets under both supervised and unsupervised settings, we establish a robust baseline and provide practical insights into the most effective combinations of data abstractions, representation models, and learning paradigms for this challenge.

cs.CR↗

arXiv2025

Read Before You Think: Mitigating LLM Comprehension Failures with Step-by-Step Reading

Large Language Models (LLMs) often fail on complex reasoning tasks due to flawed question comprehension, not just flawed logic. This paper presents a systematic investigation into these comprehension failures. Our work yields three key insights: (1) the step-by-step principle, effective for calculation, can be migrated to the reading process to enhance comprehension; (2) increasing the proportion of question-related tokens (e.g., via repetition) succeeds by refocusing attention, a mechanism that can be explicitly controlled; and (3) backward dependencies represent a core bottleneck for decoder-only models that persists even with strong methods like Chain-of-Thought. Based on these findings, we introduce the Step-by-Step Reading (SSR) family of prompts. This multi-stage approach culminates in SSR++, a method specifically engineered to deepen model comprehension by guiding it to parse questions with finer granularity, focus attention on critical tokens, and resolve backward dependencies through iterative re-contextualization. SSR++ sets a new state-of-the-art on multiple reasoning benchmarks, and our analysis confirms it works by directly mitigating semantic misunderstanding. These results demonstrate that guiding how a model reads is a powerful and efficient method for improving its reasoning ability.

cs.CL↗

arXiv2025

APFL: Analytic Personalized Federated Learning via Dual-Stream Least Squares

Personalized Federated Learning (PFL) has presented a significant challenge to deliver personalized models to individual clients through collaborative training. Existing PFL methods are often vulnerable to non-IID data, which severely hinders collective generalization and then compromises the subsequent personalization efforts. In this paper, to address this non-IID issue in PFL, we propose an Analytic Personalized Federated Learning (APFL) approach via dual-stream least squares. In our APFL, we use a foundation model as a frozen backbone for feature extraction. Subsequent to the feature extractor, we develop dual-stream analytic models to achieve both collective generalization and individual personalization. Specifically, our APFL incorporates a shared primary stream for global generalization across all clients, and a dedicated refinement stream for local personalization of each individual client. The analytical solutions of our APFL enable its ideal property of heterogeneity invariance, theoretically meaning that each personalized model remains identical regardless of how heterogeneous the data are distributed across all other clients. Empirical results across various datasets also validate the superiority of our APFL over state-of-the-art baselines, with advantages of at least 1.10%-15.45% in accuracy.

cs.LG↗

arXiv2023

A Semi-supervised Sensing Rate Learning based CMAB Scheme to Combat COVID-19 by Trustful Data Collection in the Crowd

The recruitment of trustworthy and high-quality workers is an important research issue for MCS. Previous studies either assume that the qualities of workers are known in advance, or assume that the platform knows the qualities of workers once it receives their collected data. In reality, to reduce costs and thus maximize revenue, many strategic workers do not perform their sensing tasks honestly and report fake data to the platform, which is called False data attacks. And it is very hard for the platform to evaluate the authenticity of the received data. In this paper, an incentive mechanism named Semi-supervision based Combinatorial Multi-Armed Bandit reverse Auction (SCMABA) is proposed to solve the recruitment problem of multiple unknown and strategic workers in MCS. First, we model the worker recruitment as a multi-armed bandit reverse auction problem and design an UCB-based algorithm to separate the exploration and exploitation, regarding the Sensing Rates (SRs) of recruited workers as the gain of the bandit. Next, a Semi-supervised Sensing Rate Learning (SSRL) approach is proposed to quickly and accurately obtain the workers' SRs, which consists of two phases, supervision and self-supervision. Last, SCMABA is designed organically combining the SRs acquisition mechanism with multi-armed bandit reverse auction, where supervised SR learning is used in the exploration, and the self-supervised one is used in the exploitation. We theoretically prove that our SCMABA achieves truthfulness and individual rationality and exhibits outstanding performances of the SCMABA mechanism through in-depth simulations of real-world data traces.

cs.HC↗