Searcharxiv⌕ Search

arXiv subjects

Taiqiang Wu

Publications and source records attributed to Taiqiang Wu.

At least 19 recordsLinked to original sources
arXiv2026

SafeRI: Recognition and Intervention for Token-Level Safety Intervention in Large Vision Language Models

Existing safety alignment methods for vision-language models usually modify the model behavior globally: once the safety parameters are trained or loaded, they participate in both unsafe and already-safe generations. This always-on intervention can unnecessarily perturb the model's original reasoning path and degrade general multimodal capabilities. We argue that safety alignment should be an on-demand intervention rather than a permanent modification to every decoding trajectory. To this end, we propose a streaming recognition and gated LoRA framework for intrinsic VLM safety. During autoregressive generation, a lightweight recognizer estimates whether the current pre-token generation state is safe or unsafe. Its output updates the LoRA gate for the following decoding step; otherwise, generation follows the frozen-backbone policy. The LoRA module is trained from unsafe prefixes, transition statements, and safe continuations, so that it learns to redirect unsafe generations back to safe responses after activation. Experiments across multiple safety and general-purpose benchmarks demonstrate the effectiveness of our method in post-alignment settings.

cs.CV↗
arXiv2026

OptiMAS: Automatically Optimize Multi-Agent System

Automated evolution of Multi-Agent Systems (MAS) holds significant potential for reducing the manual effort required to design and optimize LLM-based agent architectures. However, extant search-based paradigms face a fundamental trade-off, where an expanded optimization scope exacerbates evolutionary instability, while discrete branch-and-discard search isolates insights across lineages. To address these limitations, we propose a continuous, data-driven optimization paradigm built upon a unified ReAct-based infrastructure that reconciles a broad optimization scope with operational stability. Under this paradigm, we present OptiMAS, a task-agnostic agentic optimizer that leverages textual interaction trajectories and task feedback as loss signals for end-to-end MAS evolution. Equipped with a novel dual-track memory mechanism, OptiMAS sustains performance improvement over extended optimization horizons. Evaluation on four heterogeneous agentic benchmarks with three varying scale and accessibility LLM backbones, demonstrates that OptiMAS consistently achieves competitive or superior accuracy relative to both domain-specialized hand-crafted systems and existing evolutionary methods. Our work establishes a practical milestone toward robust, automated MAS evolution.

cs.MA↗
arXiv2026

SafeCap: Improving LVLM Safety with Image Captioning Reinforcement Learning

Large vision-language models (LVLMs) remain vulnerable to jailbreak attacks that exploit visual inputs to bypass safety alignment inherited from their language backbones. We propose SafeCap, a reinforcement-learning framework that aligns LVLMs through learned self-captioning. SafeCap trains a policy model to first generate a safety-relevant image caption and then produce a final answer; the caption is further optimized by whether it enables a frozen LLM to reach a safety-aligned decision. This caption-mediated objective encourages the policy to expose visual cues relevant to safe response generation rather than relying solely on direct refusal supervision. Across five multimodal safety benchmarks and six vision-utility benchmarks, SafeCap substantially improves aggregate safety performance under its intended DirectCap protocol, with gains of 3.7-19.0 points in safety average across four model settings while maintaining comparable or improved vision utility. Under controlled comparisons on matched backbones and data, SafeCap outperforms safety SFT, DPO, and SafeGRPO, demonstrating the effectiveness of caption-mediated reinforcement learning for multimodal safety alignment.

cs.CV↗
arXiv2026

Simple-OPD: Demystifying Warm-up for On-policy Distillation

On-policy distillation (OPD) trains a student on its own rollouts with token-level supervision from teacher models, but its effectiveness can depend strongly on the warm-up stage before OPD. In this paper, we demystify warm-up for OPD from both data and training perspectives. For data, we find that effective warm-up relies on teacher-compatible chain-of-thought supervision, and that even incorrect teacher rollouts can provide comparable benefits to correct ones. This suggests that warm-up primarily transfers a teacher-compatible thinking pattern rather than merely correct answers. For training, we show that low-rank adaptation (LoRA) with a near-saturation training duration better balances in-domain adaptation and out-of-distribution generalization than full-parameter SFT. Based on these findings, we propose Simple-OPD, a plug-and-play initialization method that warms up the student on teacher-generated CoT with LoRA before OPD. Experiments across diverse settings demonstrate the effectiveness and robustness of Simple-OPD.

cs.CL↗
arXiv2026

PatchINR: Patch-Based Implicit Neural Representations for Efficient and Scalable Inference

Implicit Neural Representation (INR) provides an effective approach for continuous signal modeling, but classical per-pixel inference results in quadratic growth in inference count, leading to dramatically increased computational costs in high-resolution application scenarios. To address this issue, we propose a patch-based approach that treats non-overlapping patches as fundamental processing units and predicts entire pixel patches in a single forward pass, significantly reducing the number of inference queries required. To validate the effectiveness of our approach, we propose a hardware acceleration architecture on the Field Programmable Gate Array (FPGA) platform for the INR model, which features a configurable pipeline and supports dual-precision computation. Our patch-based INR achieves comparable reconstruction quality to pixel-level INR (34.97 dB PSNR with 2 x 2 patches) while reducing inference latency by 75% with only 0.6% parameter overhead.

cs.CV↗
arXiv2026

Skim and Skip: Hierarchical Adaptive Inference for Efficient Multimodal Retrieval

Universal multimodal retrieval (UMR) increasingly adopts multimodal large language models (MLLMs) as unified embedding backbones, but their strong retrieval performance comes at substantial inference cost. Existing methods typically rely on uniformly dense inference, where all input tokens are processed through the entire model and matched using the final-layer [EOS] representation. However, this paradigm overlooks two key forms of heterogeneity in multimodal retrieval: token contributions to the final retrieval embedding are highly uneven, and different queries require markedly different amounts of inference depth. To address this, we propose Skim and Skip (SAS), a hierarchical adaptive inference framework for efficient multimodal retrieval. SAS first performs token-level evidence selection to preserve only the input information most relevant to the final retrieval embedding, and then performs depth-adaptive inference to determine whether the current representation is already sufficient for reliable matching. Experiments on 12 MMEB retrieval tasks show that SAS retains about 99% of the dense baseline's average retrieval performance while achieving up to 1.64 times end-to-end speedup and up to 66.3% FLOPs reduction.

cs.IR↗
arXiv2026

From Context-Aware to Conflict-Aware: Generalizing Contrastive Decoding for Knowledge Conflict in LLMs

When large language models generate from retrieved or augmented contexts, conflicts between external context and parametric priors remain a central reliability bottleneck. Existing contrastive decoding methods follow a \emph{context-aware} paradigm that unilaterally amplifies context over parametric priors, overwriting correct priors when the context is erroneous. We generalize this to the \textbf{conflict-aware} paradigm that dynamically allocates authority between prior and context based on conflict signals, rather than presupposing context trustworthiness. We show that the affine combination of prior and context logits yields a \textbf{power family} with an inherent \textbf{regime asymmetry}: extrapolation amplifies errors unboundedly when the prior is correct, interpolation under-corrects when the context is correct, and no static regime covers both. Existing contrastive decoding methods are instances of this family, mostly extrapolative. To evaluate both conflict directions, we propose TriState-Bench, a model-aware evaluation protocol that calibrates per-model prior knowledge to measure three conflict states: correction, resistance, and agreement. To resolve the asymmetry, we propose Adaptive Regime Routing (ARR), which routes between regimes at each step, lifting resistance EM from below 6 to 16--33 without sacrificing correction or agreement. Our code is available at https://github.com/keith-Jiang/conflict-aware-decoding.

cs.AI↗
arXiv2026

Attention Sink in Transformers: A Survey on Utilization, Interpretation, and Mitigation

As the foundational architecture of modern machine learning, Transformers have driven remarkable progress across diverse AI domains. Despite their transformative impact, a persistent challenge across various Transformers is Attention Sink (AS), in which a disproportionate amount of attention is focused on a small subset of specific yet uninformative tokens. AS complicates interpretability, significantly affecting the training and inference dynamics, and exacerbates issues such as hallucinations. In recent years, substantial research has been dedicated to understanding and harnessing AS. However, a comprehensive survey that systematically consolidates AS-related research and offers guidance for future advancements remains lacking. To address this gap, we present the first survey on AS, structured around three key dimensions that define the current research landscape: Fundamental Utilization, Mechanistic Interpretation, and Strategic Mitigation. Our work makes a pivotal contribution by highlighting the key concepts and main trends in the field, guiding researchers through the evolution of AS-related studies. We envision this survey as a valuable resource, empowering researchers to effectively manage AS within the current Transformer paradigm, while simultaneously inspiring innovative advancements for the next generation of Transformers. The paper list of this work is available at https://github.com/ZunhaiSu/Awesome-Attention-Sink.

cs.LG↗
arXiv2026

Schedule-Level Shared-Prefix Reuse for LLM RL Training

GRPO-based LLM post-training commonly samples multiple trajectories from the same prompt and then trains on the resulting group. In long-context GRPO workloads, this shared prompt-side prefix can contain retrieved passages, visual tokens, tool schemas, system instructions, or task context, while the full rollout group is still too large to pack into one training microbatch. Standard dense trainers therefore recompute the same prefix forward and backward for every trajectory. We present a schedule-level reuse mechanism that decouples prefix and suffix computation. The schedule runs prefix forward once, executes suffixes as ordinary microbatches while reading prefix K/V and accumulating prefix-side gK/gV , and then runs prefix backward once on the accumulated gradient cache. This reordered schedule is equivalent to baseline training over real arithmetic and aligns numerically within finite-precision tolerance. Because only K/V and gK/gV are hot during suffix computation, the approach offloads dormant prefix activations, integrates with TP/EP/CP/PP and DP-style placement at the execution level, and preserves aux-loss-based MoE router semantics through logical prefix-token accounting. On dense Llama3-8B, Qwen3-8B, and MoE Qwen3-MoE-30B-A3B configurations, the schedule matches optimizer updates across TP/CP/PP/EP combinations, aligns on a 100-step real GRPO actor-update trace replay, reaches up to 4.395x speedup (2.930x under a conservative compile-on comparison) as prefix ratio and GRPO group size grow, and reduces Phase-B peak HBM by up to 59.1%, extending the Llama3-8B capacity frontier from 17,920 to 29,696 total tokens.

cs.DC↗
arXiv2026

What Makes Interaction Trajectories Effective for Training Terminal Agents?

Stronger code agents are commonly assumed to be superior teachers for post-training, yet this assumption remains poorly disentangled from task difficulty, harness design, and student capacity. We investigate this pedagogical link using Terminal-Lego, a scalable pipeline that transforms multi-domain real-world issues into environment-verified agentic tasks. Surprisingly, standalone performance does not dictate teaching efficacy: while Claude Opus 4.6 achieves higher scores on Terminal-Bench 2.0, students fine-tuned on trajectories from DeepSeek-V3.2, a lower-scoring agent, exhibit significantly stronger generalization. We attribute this "pedagogical paradox" to Environment-Grounded Supervision (EGS): trajectories that explicitly expose inspect-act-verify behaviors through harness-visible interactions allow students to internalize robust problem-solving routines rather than fragile action sequences. Scaling analysis reveals exceptional data efficiency: with only 15.3k Terminal-Lego trajectories, for example, Qwen3-32B achieves a 24.3% score on Terminal-Bench 2.0, rivaling previous SOTA performance established with over 30x the data volume. Our results suggest that the frontier of agent post-training lies beyond mere outcome-matching, shifting the focus toward "Harness Engineering", where the systematic design of environment-grounded interaction structures serves as the primary catalyst for reproducible and generalizable agentic intelligence.

cs.AI↗
arXiv2026

BPDQ: Bit-Plane Decomposition Quantization on a Variable Grid for Large Language Models

Large language model inference is often bounded by memory footprint and bandwidth in resource-constrained deployments, making quantization fundamental to efficient serving. While post-training quantization (PTQ) maintains high fidelity at 4-bit, it deteriorates at 2-3 bits. In essence, existing methods enforce a shape-invariant quantization grid (e.g., the fixed uniform intervals of UINT2) for each group, severely restricting the feasible set for error minimization. To address this, we propose Bit-Plane Decomposition Quantization (BPDQ), which constructs a variable quantization grid via bit-planes and scalar coefficients, and iteratively refines them using second-order information while progressively compensating for quantization errors to minimize output discrepancy. In the 2-bit regime, BPDQ enables serving Qwen2.5-72B on a single RTX 3090 with 83.85\% GSM8K accuracy (vs. 90.83\% at 16-bit). Moreover, we theoretically show that the variable grid expands the feasible set, and that the quantization process consistently aligns with the optimization objective in Hessian-induced geometry. The code is available at https://github.com/KingdalfGoodman/BPDQ.

cs.LG↗
arXiv2026

ROMER: Expert Replacement and Router Calibration for Robust MoE LLMs on Analog Compute-in-Memory Systems

Large language models (LLMs) with mixture-of-experts (MoE) architectures achieve remarkable scalability by sparsely activating a subset of experts per token, yet their frequent expert switching creates memory bandwidth bottlenecks that compute-in-memory (CIM) architectures are well-suited to mitigate. However, analog CIM systems suffer from inherent hardware imperfections that perturb stored weights, and its negative impact on MoE-based LLMs in noisy CIM environments remains unexplored. In this work, we present the first systematic investigation of MoE-based LLMs under noise model calibrated with real chip measurements, revealing that hardware noise critically disrupts expert load balance and renders clean-trained routing decisions consistently suboptimal. Based on these findings, we propose ROMER, a post-training calibration framework that (1) replaces underactivated experts with high-frequency ones to restore load balance, and (2) recalibrates router logits via percentile-based normalization to stabilize routing under noise. Extensive experiments across multiple benchmarks demonstrate that ROMER achieves up to 58.6\%, 58.8\%, and 59.8\% reduction in perplexity under real-chip noise conditions for DeepSeek-MoE, Qwen-MoE, and OLMoE, respectively, establishing its effectiveness and generalizability across diverse MoE architectures.

cs.LG↗
arXiv2026

ProFit: Leveraging High-Value Signals in SFT via Probability-Guided Token Selection

Supervised fine-tuning (SFT) is a fundamental post-training strategy to align Large Language Models (LLMs) with human intent. However, traditional SFT often ignores the one-to-many nature of language by forcing alignment with a single reference answer, leading to the model overfitting to non-core expressions. Although our empirical analysis suggests that introducing multiple reference answers can mitigate this issue, the prohibitive data and computational costs necessitate a strategic shift: prioritizing the mitigation of single-reference overfitting over the costly pursuit of answer diversity. To achieve this, we reveal the intrinsic connection between token probability and semantic importance: high-probability tokens carry the core logical framework, while low-probability tokens are mostly replaceable expressions. Based on this insight, we propose ProFit, which selectively masks low-probability tokens to prevent surface-level overfitting. Extensive experiments confirm that ProFit consistently outperforms traditional SFT baselines on general reasoning and mathematical benchmarks.

cs.CL↗
arXiv2026

The Art of Efficient Reasoning: Data, Reward, and Optimization

Large Language Models (LLMs) consistently benefit from scaled Chain-of-Thought (CoT) reasoning, but also suffer from heavy computational overhead. To address this issue, efficient reasoning aims to incentivize short yet accurate thinking trajectories, typically through reward shaping with Reinforcement Learning (RL). In this paper, we systematically investigate the mechanics of efficient reasoning for LLMs. For comprehensive evaluation, we advocate for more fine-grained metrics, including length distribution conditioned on correctness and performance across a wide spectrum of token budgets ranging from 2k to 32k. First, we reveal that the training process follows a two-stage paradigm: length adaptation and reasoning refinement. Through extensive experiments (about 0.2 million GPU hours) in a unified protocol, we deconstruct training prompts and rollouts, reward shaping, and optimization strategies. A central finding is to maintain a sufficient density of positive reward signals and avoid the short-is-correct trap. Moreover, the learned length bias generalizes across domains and difficulty levels. We distill these findings into valuable insights and practical guidelines, and validate them across the Qwen3 models ranging from 0.6B to 30B, demonstrating the robustness and generalization. Weights are available at https://wutaiqiang.github.io/project/Art

cs.CL↗
arXiv2026

Can We Trust LLMs on Memristors? Diving into Reasoning Ability under Non-Ideality

Memristor-based analog compute-in-memory (CIM) architectures provide a promising substrate for the efficient deployment of Large Language Models (LLMs), owing to superior energy efficiency and computational density. However, these architectures suffer from precision issues caused by intrinsic non-idealities of memristors. In this paper, we first conduct a comprehensive investigation into the impact of such typical non-idealities on LLM reasoning. Empirical results indicate that reasoning capability decreases significantly but varies for distinct benchmarks. Subsequently, we systematically appraise three training-free strategies, including thinking mode, in-context learning, and module redundancy. We thus summarize valuable guidelines, i.e., shallow layer redundancy is particularly effective for improving robustness, thinking mode performs better under low noise levels but degrades at higher noise, and in-context learning reduces output length with a slight performance trade-off. Our findings offer new insights into LLM reasoning under non-ideality and practical strategies to improve robustness.

cs.CL↗
arXiv2026

A Survey of Reasoning in Autonomous Driving Systems: Open Challenges and Emerging Paradigms

The development of high-level autonomous driving (AD) is shifting from perception-centric limitations to a more fundamental bottleneck, namely, a deficit in robust and generalizable reasoning. Although current AD systems manage structured environments, they consistently falter in long-tail scenarios and complex social interactions that require human-like judgment. Meanwhile, the advent of large language and multimodal models (LLMs and MLLMs) presents a transformative opportunity to integrate a powerful cognitive engine into AD systems, moving beyond pattern matching toward genuine comprehension. However, a systematic framework to guide this integration is critically lacking. To bridge this gap, we provide a comprehensive review of this emerging field and argue that reasoning should be elevated from a modular component to the system's cognitive core. Specifically, we first propose a novel Cognitive Hierarchy to decompose the monolithic driving task according to its cognitive and interactive complexity. Building on this, we further derive and systematize seven core reasoning challenges, such as the responsiveness-reasoning trade-off and social-game reasoning. Furthermore, we conduct a dual-perspective review of the state-of-the-art, analyzing both system-centric approaches to architecting intelligent agents and evaluation-centric practices for their validation. Our analysis reveals a clear trend toward holistic and interpretable "glass-box" agents. In conclusion, we identify a fundamental and unresolved tension between the high-latency, deliberative nature of LLM-based reasoning and the millisecond-scale, safety-critical demands of vehicle control. For future work, a primary objective is to bridge the symbolic-to-physical gap by developing verifiable neuro-symbolic architectures, robust reasoning under uncertainty, and scalable models for implicit social negotiation.

cs.AI↗
arXiv2026

HaLoRA: Hardware-aware Low-Rank Adaptation for Large Language Models Based on Hybrid Compute-in-Memory Architecture

Low-rank adaptation (LoRA) is a predominant parameter-efficient finetuning method for adapting large language models (LLMs) to downstream tasks. Meanwhile, Compute-in-Memory (CIM) architectures demonstrate superior energy efficiency due to their array-level parallel in-memory computing designs. In this paper, we propose deploying the LoRA-finetuned LLMs on the hybrid CIM architecture (i.e., pretrained weights onto energy-efficient Resistive Random-Access Memory (RRAM) and LoRA branches onto noise-free Static Random-Access Memory (SRAM)), reducing the energy cost to about 3\% compared to the Nvidia A100 GPU. However, the inherent noise of RRAM on the saved weights leads to performance degradation, simultaneously. To address this issue, we design a novel Hardware-aware Low-rank Adaptation (HaLoRA) method. The key insight is to train a LoRA branch that is robust toward such noise and then deploy it on noise-free SRAM, while the extra cost is negligible since the parameters of LoRAs are much fewer than pretrained weights (e.g., 0.15\% for LLaMA-3.2 1B model). To improve the robustness towards the noise, we theoretically analyze the gap between the optimization trajectories of the LoRA branch under both ideal and noisy conditions and further design an extra loss to minimize the upper bound of this gap. Therefore, we can enjoy both energy efficiency and accuracy during inference. Experiments finetuning the Qwen and LLaMA series demonstrate the effectiveness of HaLoRA across multiple reasoning tasks, achieving up to \textbf{22.7} improvement in average score while maintaining robustness at various noise types and noise levels.

cs.CL↗
arXiv2026

SwingArena: Competitive Programming Arena for Long-context GitHub Issue Solving

We present SwingArena, a competitive evaluation framework for Large Language Models (LLMs) that closely mirrors real-world software development workflows. Unlike traditional static benchmarks, SwingArena models the collaborative process of software iteration by pairing LLMs as submitters, who generate patches, and reviewers, who create test cases and verify the patches through continuous integration (CI) pipelines. To support these interactive evaluations, we introduce a retrieval-augmented code generation (RACG) module that efficiently handles long-context challenges by providing syntactically and semantically relevant code snippets from large codebases, supporting multiple programming languages (C++, Python, Rust, and Go). This enables the framework to scale across diverse tasks and contexts while respecting token limitations. Our experiments, using over 400 high-quality real-world GitHub issues selected from a pool of 2,300 issues, show that models like GPT-4o excel at aggressive patch generation, whereas DeepSeek and Gemini prioritize correctness in CI validation. SwingArena presents a scalable and extensible methodology for evaluating LLMs in realistic, CI-driven software development settings. More details are available on our project page: swing-bench.github.io

cs.CL↗