Search arXiv⌕ Search

arXiv subjects

Li Zhang

Publications and source records attributed to Li Zhang.

At least 91 records · Page 5Linked to original sources
arXiv2026

Leveraging Machine Unlearning for Cost-Efficient Preference Alignment

Despite advances in Preference Alignment (PA) for Large Language Models (LLMs), mainstream methods like reinforcement learning with human feedback face notable challenges. These approaches require high-quality datasets of positive preference examples, which are costly to obtain and computationally intensive. The LLM unlearning technique presents a promising alternative by directly removing the influence of negative examples. However, current research has primarily focused on empirical validation, lacking systematic quantitative analysis. To bridge this gap, we propose a framework linking PA with LLM unlearning. Through bi-level optimization, we first quantify how unlearning specific negative examples impacts PA performance. Our analysis reveals that these effects vary substantially across negative examples. Building on this insight, we pose a crucial question: how can we optimally select and weight negative examples for unlearning to maximize PA performance? To answer this, we propose Unlearning to Align (U2A), which leverages bi-level optimization to efficiently select and unlearn examples for optimal PA performance. We validate the proposed method through extensive experiments, with results confirming its effectiveness. Our code is available at https://github.com/muyiahhh/U2A.

cs.LG↗
arXiv2026

Generative Model Unlearning: A Survey through Target Events, Unlearning Operators, and Evaluation Protocols

With the rapid advancement of generative models, privacy, copyright, safety, and reliability risks have attracted growing attention. To mitigate these risks, machine unlearning has been increasingly adapted from traditional classification models to generative settings. Despite notable progress, existing studies remain fragmented in their target definitions, unlearning mechanisms, and evaluation protocols, making objective comparison difficult across models, modalities, and applications. Moreover, modality-specific surveys often overlook the shared structure of Generative Model Unlearning (GenMU). To address this gap, we provide a comprehensive review of GenMU and formulate it as target-constrained distributional projection: given a target event, an unlearning operator transforms the generative distribution to suppress target-related outputs while preserving useful behavior and controlling operator cost. Under this framework, an unlearning request is specified by a target event, implemented through an unlearning operator, and assessed by empirical evidence over target suppression, distribution preservation, and operator cost. We further reorganize existing GenMU studies under this view. From this perspective, we provide the first explicit and unified account of how GenMU connects to mainstream applications, including privacy protection, copyright and style protection, safety alignment, hallucination mitigation, and deployment defense. Finally, we identify key open problems and future directions toward reliable, scalable, robust, and auditable GenMU. We consistently maintain the related open-source materials at https://github.com/caxLee/GenMU-Survey.

cs.LG↗
arXiv2026

CangjieBench: Benchmarking LLMs on a Low-Resource General-Purpose Programming Language

Large Language Models excel in high-resource programming languages but struggle with low-resource ones. Existing research related to low-resource programming languages primarily focuses on Domain-Specific Languages (DSLs), leaving general-purpose languages that suffer from data scarcity underexplored. To address this gap, we introduce CangjieBench, a contamination-free benchmark for Cangjie, a representative low-resource general-purpose language. The benchmark comprises 248 high-quality samples manually translated from HumanEval and ClassEval, covering both Text-to-Code and Code-to-Code tasks. We conduct a systematic evaluation of diverse LLMs under four settings: Direct Generation, Syntax-Constrained Generation, Retrieval-Augmented Generation (RAG), and Agent. Experiments reveal that Direct Generation performs poorly, whereas Syntax-Constrained Generation offers the best trade-off between accuracy and computational cost. Agent achieve state-of-the-art accuracy but incur high token consumption. Furthermore, we observe that Code-to-Code translation often underperforms Text-to-Code generation, suggesting a negative transfer phenomenon where models overfit to the source language patterns. We hope that our work will offer valuable insights into LLM generalization to unseen and low-resource programming languages. Our code and data are available at https://github.com/cjhCoder7/CangjieBench.

cs.SE↗
arXiv2026

TangPoetryBench: A Multi-Dimensional Benchmark and Rubric-Conditioned Evaluator for Poetry-to-Image Generation

Text-to-image (T2I) models are increasingly asked to illustrate literary and cultural content, yet we cannot measure how well an image renders the meaning of a poem. The task is many-sided: a good illustration must be visually sound, faithful to the poem's imagery and scene, culturally and stylistically apt, free of spurious text, and true to its emotion, and its deepest requirements, imagery and especially implicit emotion, are never stated in the words. Existing metrics (CLIPScore, BLIPScore, VQAScore) reward literal text-image correspondence and so cannot tell whether an illustration succeeds, let alone why, or even separate the best model from the worst. We introduce TangPoetryBench, a multi-dimensional benchmark of 1,280 images (320 classical Chinese Tang poems x 4 state-of-the-art T2I models) with quality-controlled human annotations across ten dimensions. Analyzing this data, we reveal the shared and model-specific strengths and weaknesses of current T2I models, including their ability to evoke a poem's implicit emotion. We further introduce PoemAutoEvaluator (PAE), an open, rubric-conditioned evaluator that reaches parity with a strong proprietary judge (Claude), generalizes to an unseen generator and a second poetic tradition (Song Ci), and lets the benchmark scale to new images without fresh human annotation. We release the benchmark, annotations, and evaluator.

cs.CV↗
arXiv2026

Minute-Scale Training for Microrobot Navigation

Microrobots hold significant potential for various applications, where targeted navigation is a basic requirement. Deep reinforcement learning (DRL) has recently emerged as a powerful paradigm for fully autonomous microrobot navigation. Yet, current DRL-based approaches pay limited attention to learning efficiency and effectiveness, requiring hours to days for model training. Consequently, this impedes both rapid practical deployment and parameter optimization. To address these challenges, we present a learning framework that enables effective microrobot navigation policies to be trained within minutes. In the proposed framework, we develop a fully vectorized simulator with more than 10,000 artificial vascular environments, parallelizing dynamics, LiDAR-inspired perception, and feasibility checks across thousands of environments to achieve roughly 190,000 transitions per second. To achieve effectiveness in fast training, we propose a task-shaping-regularization (TSR) reward framework. The TSR framework accelerates convergence, improves final performance, reduces action variation by at least 33.7%, and increases obstacle clearance by at least 2.1% across all evaluated scenarios. Results show that the proposed learning framework reduces training time to under 10 minutes, while supporting zero-shot deployment across distinct microrobot types and navigation scenarios. Collectively, this framework can substantially shorten the design loop and accelerate the deployment of autonomous microrobots.

cs.RO↗
arXiv2026

Long-Horizon Agent Trajectory Attribution: A Unified Benchmark and Fine-Grained Annotation Framework

Large language model (LLM) agents increasingly operate through long-horizon trajectories involving user instructions, tool use, external observations, and memory. Existing benchmarks primarily evaluate behavioral outcomes but provide limited support for fine-grained attribution analysis. We introduce trajectory attribution and develop a benchmark and annotation framework for this task. The benchmark organizes heterogeneous trajectories under a unified component schema and provides annotations of the primary attribution component, together with attack and execution chains where applicable. Instantiating the benchmark with trajectories from AgentDojo and the Stage and Canary settings of Agent3Sigma yields more than 1,300 annotated trajectories covering task-aligned actions, unsafe actions, and safety refusals. The benchmark defines two evaluation tasks, primary attribution localization and attribution-chain recovery, and provides reference baselines based on incremental trajectory contribution and component-level leave-one-out perturbation. It captures diverse attribution settings, including local and long-range attribution as well as structured attribution chains. Reference baseline results exhibit substantial performance differences across these settings, providing an initial characterization of the benchmark's attribution challenges. Beyond this initial instantiation, we release a reusable annotation skill that enables trajectories generated by new agent models to be standardized, annotated, and evaluated under the same framework. Project resources and future releases are available at https://github.com/chenjing-2024/agent-trajectory-attribution.

cs.AI↗
arXiv2026

Controlled Memory Interference in Continual LLM Agents

Long-term memory enables AI agents to maintain continuity across sessions, personalize behavior, and evolve through accumulated experience. Yet memory evolution is not simply a process of storing more information: new experiences may reinforce, revise, or interfere with existing memory states. Existing systems mainly emphasize memory construction and relevance-based retrieval, but several memories may remain simultaneously relevant while differing in state, temporal validity, or authority. We introduce Controlled Memory Interference (CMI), a controlled diagnostic and data-generation framework for studying how agent memory evolves under different memory relationships. Across controlled memory evolution, benign accumulation has limited effects, whereas relationship-specific interference sharply suppresses update plasticity with little stability gain, either by blocking target-memory exposure or by disrupting its downstream use. Lexical and Dense retrieval exhibit distinct interference pathways, while poisoning is more sensitive to update-authority cues than to recency alone. Beyond diagnosis, CMI provides targeted examples for interference-aware memory learning, improving the distinction between valid updates and interference-inducing memories while preserving performance on original memory tasks. These findings show that memory evolution is shaped not only by memory scale, but also by interactions among accumulated experiences. More broadly, memory interference emerges as an important factor for reliable continual agent memory systems.

cs.AI↗
arXiv2026

Reinforcing Action Policies by Prophesying

Vision-Language-Action (VLA) policies excel in aligning language, perception, and robot control. However, most VLAs are trained purely by imitation, which overfits to demonstrations, and is brittle under distribution shift. Reinforcement learning (RL) directly optimizes task reward and thus addresses this misalignment, but real-robot interaction is expensive and conventional simulators are hard to engineer and transfer. We address both data efficiency and optimization stability in VLA post-training via a learned world model and an RL procedure tailored to flow-based action heads. Specifically, we first introduce Prophet, a unified action-to-video robot world model pretrained on large-scale, heterogeneous robot data to learn reusable action-outcome dynamics and then few-shot adapted to new robots, objects, and environments, yielding a rollout-ready simulator. Upon Prophet, we reinforce action policies with our proposed FlowScale, which couples Flow-GRPO with intrinsic stepwise reweighting to stabilize gradients. Together, our solution provides a practical, data- and compute-efficient path to VLA post-training. Experiments show 5-17% success gains on public benchmarks and 24-30% on real robots across diverse VLA backbones.

cs.RO↗
arXiv2026

Answer Presence Drives RAG Rewriting Gains

Retrieval-augmented QA pipelines often route retrieved passages through an LLM \emph{rewriter} before a smaller reader, lifting F1 by tens of points on multi-hop benchmarks; this gain is typically credited to improved evidence quality. We ask whether that lift is causally driven by the gold answer string appearing in the rewritten context rather than by curation per se, using a controlled intervention audit. For each rewritten context we re-run the reader after one of four controlled edits to the compile output: removing the gold answer span, replacing a length-matched random non-answer span (placebo), or injecting the gold into rewrites where it was absent (at the prefix or at a midpoint sentence boundary). Across twelve completed (cell, baseline) intervention runs spanning three reader families (Qwen2.5-7B, Qwen3.5-35B, GLM-4.7), two datasets (HotpotQA, 2WikiMultihopQA), and three compiler arrangements (MA-only, MB-only, MA$+$verify), removing the gold answer drops reader F1 by $28$ to $64$ points beyond the length-matched placebo on paired \texttt{answer-in-compile} strata, and prepending the gold into rewrites that lacked it raises F1 by $+0.7$ to $+9.7$ points in $10$ of $12$ (cell, baseline) combinations. A companion five-sentinel audit shows the conventional single-\texttt{[MASK]} probe is itself sentinel-fragile: on 2Wiki it reports a $+4.12$~F1 ``non-leakage residual'' that flips to $-3.33$ to $-7.81$~F1 under four alternative sentinels and fails an equivalence test for three of those four ($1/4$~pass). We do not propose a new rewriter or mitigation; we release the intervention runner and the sentinel panel so that other rewriter-gain claims can be tested against the same standard.

cs.AI↗
arXiv2026

From Patches to Evidence Balls: Class-Conditioned Evidence Retrieval for Few-Shot Whole Slide Image Classification

Whole slide image (WSI) classification is an evidence-driven task, where diagnostic cues are often sparse, spatially organized, and class-dependent. Existing MIL and vision-language methods aggregate a large pool of patch features into a single global slide representation. Under few-shot supervision, limited slide-level labels make it difficult to learn a reliable aggregation mechanism that organizes sparse local cues into compact and coherent diagnostic evidence. Moreover, a shared slide representation compresses evidence supporting a candidate class and its alternatives into the same feature, limiting class-specific reasoning and interpretability. To address these issues, we propose EviBall, a class-conditioned evidence retrieval framework for few-shot WSI classification. EviBall organizes local patches into Evidence Balls through semantic-spatial assignment and center refinement, yielding compact and spatially coherent evidence units under weak supervision. It then uses task-specific class queries, including language-guided queries for morphology-oriented tasks and molecular-guided queries for molecular endpoint prediction, to retrieve supporting evidence balls and produce class-conditioned evidence representations for direct class-wise prediction. By introducing structured evidence units and task-relevant semantic guidance, EviBall reduces the reliance on learning an unconstrained global aggregation mechanism from scarce slide-level labels. It therefore reformulates few-shot WSI classification as structured evidence retrieval and competition among candidate classes. Extensive experiments across four morphology-oriented and molecular endpoint WSI tasks demonstrate that EviBall consistently outperforms conventional and vision-language MIL baselines under diverse few-shot settings, while providing spatially localized and class-specific evidence for each prediction.

cs.CV↗
arXiv2026

PhySR: Physics-Informed Neural Network for Super-Resolution Reconstruction in Radio Synthesis Imaging

Radio telescope arrays are constrained by the number of antennas and baseline distribution, resulting in incomplete spatial-frequency sampling, limited image resolution, and blurring, distortion, and loss of small-scale structures caused by coupling between the primary and synthesized beams. Existing general-purpose model-driven methods remove observational effects sequentially and may accumulate errors, but cannot directly address limited imaging resolution, while data-driven methods lack explicit physical constraints. We propose PhySR, an end-to-end physics-informed neural network that combines a U-Net backbone, dynamic cascaded upsampling, a multiscale feature residual module, and a differentiable physical forward model incorporating the primary beam response, PSF convolution, and scale mapping. PhySR directly reconstructs high-resolution images from low-resolution dirty images without high-resolution labels while maintaining observation-domain consistency. Experiments on simulated SKA-Mid data show that, for 4x super-resolution, PhySR achieves a PSNR of 44.65 dB, an SSIM of 0.9940, and an RMSE of 0.0065. Compared with existing general-purpose methods, PSNR and SSIM improve by approximately 13.23 dB and 0.3760, respectively. Compared with mainstream deep learning models, PSNR and SSIM improve by 6.50 dB and 0.0682, while RMSE decreases by 0.0069. PhySR also remains stable for 2x and 8x super-resolution and achieves low observation-domain consistency errors, demonstrating advantages in coupling-effect removal, small-scale structure recovery, and physical consistency.

astro-ph.IM↗
arXiv2026

Beyond "What to Retrieve": Uncertainty in Retrieval-Augmented Code Generation

Repository-level code generation relies on heterogeneous evidence whose relevance, compatibility, and completeness are inherently uncertain. Similar-code examples, repository context, and project-specific APIs may provide complementary information, but can also introduce noisy, redundant, or conflicting signals. Existing retrieval-augmented approaches primarily optimize retrieval relevance without explicitly modeling how uncertainty in retrieved evidence affects downstream generation. We introduce OpenCoder, an uncertainty-aware framework that estimates source-specific uncertainty, uses it to filter and rank heterogeneous evidence, and guides generation, verification, and repair. A factorial analysis over API knowledge, repository context, and similar-code evidence reveals no universal additive source ranking; instead, significant cross-source interactions depend on the accompanying evidence and LLM backend. On an expanded 32-task RepoExec-inline evaluation, OpenCoder improves GPT selected-output correctness over Baseline RAG from 56.25\% to 78.13\%. However, it matches a verification-and-repair control, and the corresponding Gemini improvement is not statistically supported, indicating backend-dependent benefits. Target-aware API refinement also substantially improves API-set retrieval. These findings support treating uncertainty as an actionable control signal for repository-level retrieval, verification, and repair.

cs.SE↗
arXiv2026

Beyond Self-Knowledge: Propagating Uncertainty Across Reasoning and Retrieval in LLMs

Retrieval-augmented generation improves knowledge-intensive question answering, but indiscriminate retrieval can introduce irrelevant evidence and unnecessary computation. We investigate whether verbalized confidence from black-box language models can serve as an actionable signal for retrieval routing. Our method, BeyondUncertainty, first elicits a structured provisional answer and confidence estimate, then applies a model-specific threshold selected on held-out validation data and frozen before test evaluation. Low-confidence questions receive top-5 TF-IDF retrieval followed by a second answer call, whereas high-confidence questions return the provisional answer directly. We evaluate 27,000 policy instances across six QA benchmarks, three model families, and three retrieval policies. BeyondUncertainty achieves 0.483 mean token-level F1, compared with 0.467 for always retrieval and 0.401 for no retrieval, while reducing retrieved passages by 20.4\% relative to always retrieval. When matched on the number of questions routed to retrieval within each dataset-model cell, it outperforms a post-hoc random allocation in 17 of 18 settings, with an average gain of 0.024 F1. Although poorly calibrated as an absolute probability, probe uncertainty modestly predicts question-level retrieval benefit (AUROC = 0.628). However, the additional probe increases total token usage by 28.2\%, revealing a trade-off between more selective evidence acquisition and end-to-end token efficiency.

cs.IR↗
arXiv2026

CodeSpec: Dual Executable Specifications for Agentic Long-Horizon Feature Development

LLM-based code agents have advanced repository-level software development through iterative interaction with codebases and tools. However, feature development requires integrating new behaviors into existing architectures through coherent cross-component functional chains. Existing agents typically derive such chains through free-form reasoning, often producing unreliable feature designs with incomplete functional chains. Moreover, textual designs are difficult to verify and enforce, making it challenging to maintain design-implementation consistency throughout long-horizon development. We propose CodeSpec, a dual executable specification method for repository-level feature development. It builds reliable functional chains from evidence pairing sub-requirement semantics with repository architectures, then compiles them into complementary architecture and behavior specifications that check chain completeness and correctness while preserving design-implementation consistency over long interactions. On FeatureBench, which targets feature development in existing repositories, CodeSpec achieves 70.7%, 55.0%, and 49.9% pass rates under DeepSeek-V4-Pro, outperforming representative baselines such as Claude Code. Results on the repository generation benchmark NL2Repo-Bench further demonstrate its generalizability.

cs.SE↗
arXiv2026

MRCoder: An Efficient Context Selecting Approach for Repository-Level Code Generation

Large language models (LLMs) have demonstrated strong capabilities in code generation. However, repository-level code generation remains challenging, as it requires effectively identifying and utilizing repository-specific context. While retrieval-augmented generation (RAG) incorporates relevant code snippets, it often introduces redundant context that interferes with the LLM's ability to utilize relevant information, leading to degraded generation quality and increased computational cost. Moreover, existing context selection and compression methods struggle to balance efficiency and quality, either introducing additional computational overhead or failing to effectively select valid context. In this paper, we propose MRCoder, an efficient context selection framework that improves both the effectiveness and efficiency of repository-level code generation. MRCoder adopts a Map-Reduce paradigm: in the Map Phase, a lightweight draft model generates drafts over partitioned contexts, and Structure-Aware Draft-Guided Selection (SADGS) selects informative contexts based on drafts through API consistency and logical similarity; in the Reduce Phase, the refined contexts are aggregated for final generation, with a parallel verification strategy further accelerating decoding. We evaluate MRCoder on two widely used repository-level code generation benchmarks, CoderEval and DevEval, using Qwen2.5-Coder and DeepSeek-Coder as backbone LLMs. Experimental results show that MRCoder improves code generation accuracy over strong baselines while reducing token consumption by 30 to 50% and inference time by up to 52%. These results demonstrate that our proposed structured and draft-guided context selection strategy is crucial for improving both the quality and efficiency of repository-level code generation

cs.SE↗
arXiv2026

Gemma 4 Technical Report

We introduce Gemma 4, a new generation of open-weight, natively multimodal language models in the Gemma model family. Designed to advance compute efficiency and reasoning, the Gemma 4 model suite features dense and Mixture-of-Experts architectures, ranging from 2.3B to 31B parameters. Alongside improved vision and audio encoders for all model sizes, we propose a unified, encoder-free architecture for our 12B model, which ingests raw audio and image patches. Furthermore, we integrate a thinking mode, enabling Gemma models to generate reasoning traces prior to responding. We improve inference speed, memory, and compute efficiency, as well as long-context abilities through critical design choices. Gemma 4 establishes a leap in performance across STEM, multimodal, and long-context benchmarks, and rivals larger, frontier open models in human-rated tasks.

cs.CL↗
arXiv2026

Inter-Reflective Gaussian Splatting for Robust and Efficient Inverse Rendering

Faithful inverse rendering requires visibility and indirect radiance to explain secondary illumination and inter-reflection, yet rasterization-oriented Gaussian representations do not naturally support the secondary-ray queries needed to recover them. We present IRGS++ (Inter-Reflective Gaussian Splatting), a unified robust and efficient Gaussian inverse rendering framework. During transport-aware optimization, IRGS++ employs differentiable 2D Gaussian ray tracing on surface-oriented Gaussian primitives to query visibility and indirect radiance on the fly and evaluate the full rendering equation for inter-reflective transport. This physical core makes Gaussian inverse rendering physically grounded beyond rasterized appearance modeling. To make this backbone useful beyond low-gloss dielectric scenes, the framework incorporates metallic-aware material modeling and robust reflective initialization for glossy, specular, and metallic materials. To make it practical, multiple importance sampling and denoising stabilize finite-sample rendering, while mesh-based secondary-attribute queries reduce the cost of relighting under novel illumination. Quantitative evaluations on low-gloss and glossy benchmarks show improved decomposition and relighting quality together with favorable quality--speed trade-offs under the reported configurations, while real-world studies illustrate plausible relighting under novel illumination.

cs.GR↗
arXiv2026

DSBench: A Comprehensive Benchmark for Evaluating External and In-Cabin Risks

Vision-Language Models (VLMs) show great promise for autonomous driving, but their suitability for safety-critical scenarios is largely unexplored, raising safety concerns. This issue arises from the lack of comprehensive benchmarks that assess both external environmental risks and in-cabin driving behavior safety simultaneously. To bridge this critical gap, we introduce DSBench, the first comprehensive Driving Safety Benchmark designed to assess a VLM's awareness of various safety risks in a unified manner. DSBench encompasses two major categories: external environmental risks and in-cabin driving behavior safety, divided into 10 key categories and a total of 28 sub-categories. This comprehensive evaluation covers a wide range of scenarios, ensuring a thorough assessment of VLMs' performance in safety-critical contexts. Extensive evaluations across various mainstream open-source and closed-source VLMs reveal significant performance degradation under complex safety-critical situations, highlighting urgent safety concerns. To address this, we constructed a large dataset of 98K instances focused on in-cabin and external safety scenarios, showing that fine-tuning on this dataset significantly enhances the safety performance of existing VLMs and paves the way for advancing autonomous driving technology. The benchmark toolkit, code, and model checkpoints will be publicly accessible.

cs.RO↗