Searcharxiv⌕ Search

arXiv subjects

Yusen Zhang

Publications and source records attributed to Yusen Zhang.

At least 19 recordsLinked to original sources
arXiv2026

CIRCE Divines LLM Task Complexity from Constraint Structure

Estimating whether a problem is too complex for a large language model (LLM) before the model attempts it can guide model routing, how to allocate test-time compute, and suggest the need for new tools to aid the agent. While existing estimators rely on the problem text, LLM-annotated task difficulty labels, or signals obtained after a model attempts to solve the problem, such as its hidden states, none explicitly leverage the constraint structure inherent in the problem: the variables to assign and how the constraints over them interact. We introduce CIRCE, which extracts a problem's structure by using a cheap LLM to construct a Constraint Intermediate Representation that deterministically compiles into the constraint graph. Across seven reasoning benchmarks (including SAT, logic, and math) and four LLMs, even a randomly initialized GNN encoder improves predictions of LLM response correctness, reducing negative log-likelihood from 0.465 to 0.423 and increasing AUROC from 0.691 to 0.727 relative to the strongest baseline for each metric.

cs.AI↗
arXiv2026

VISTA: A Versatile Interactive User Simulation Toolkit for Agent Evaluation

Evaluation remains a critical bottleneck for interactive agent development. Existing evaluation methods often rely on static benchmarks, which fail to capture the dynamic, multi-step nature of agentic behavior and struggle to expose meaningful failure modes. While user-simulation-based evaluation offers a promising alternative, existing simulation frameworks suffer from two major limitations. First, they provide limited mechanisms for evaluating the quality and comprehensiveness of simulated interactions, making it difficult to assess whether a simulator sufficiently explores an agent's capabilities and failure modes. Second, most frameworks are restricted to either UI-only actions or API-only actions, limiting their ability to model the full range of realistic user behaviors. To address these limitations, we propose VISTA, a Versatile Interactive user Simulation Toolkit for Agent evaluation. Our toolkit includes a suite of six metrics for measuring the realism, capability coverage, and interaction effectiveness of simulated interactions. In addition, we develop a hybrid user simulator that integrates both UI-based interactions and API-based interactions, enabling more realistic and comprehensive evaluation across diverse interactive environments. We evaluate VISTA in e-commerce shopping and education customer service settings and demonstrate that it produces more realistic and comprehensive evaluations than existing methods.

cs.CL↗
arXiv2026

AgentWorld: Benchmarking Long-Horizon Collaboration of Multi-agent LLMs

Existing multi-agent benchmarks primarily test in competitive settings, short-horizon interactions under 20 steps, or simply aggregate individual performance, failing to isolate and highlight genuine collaboration capabilities of LLM-based agents. We introduce AgentWorld, a benchmark of 100 human-annotated tasks (with 100 augmented variants) for evaluating long-horizon, multi-agent collaboration. Tasks span 50+ interaction rounds across a rich MMORPG sandbox and require 3-20 agents with asymmetric roles and abilities to coordinate through communication, joint planning, and resource sharing under a blackbox setting where each agent acts independently without access to others' internal states. To quantify collaboration effectiveness in addition to conventional binary task success, we propose Causal Collaboration Effectiveness (CCE), a graph-based metric that traces causal dependencies between agent actions and measures what fraction of a team's effort actually contributed to the outcome. Experiments with Gemini 3 Flash, Claude Haiku 4.5, GPT-5 Mini, and DeepSeek R1-70B show that even the best model achieves only 52.0% task success, with systematic failure modes including communication breakdowns, role confusion, and inability to maintain shared plans across rounds. AgentWorld is fully open-source.

cs.MA↗
arXiv2026

Receiver-Conditioned Latent Communication gives 94% CacheBack

Multi-agent systems distribute large contexts across agents that communicate to solve a task. Text messages are compact but require decoding and may omit evidence the receiving agent needs. Recent latent communication instead transfers KV caches. This avoids text generation and can improve accuracy and latency. However, a full KV cache grows linearly with both the context an individual agent processes, and the number of agents that coordinate together. This raises memory and context costs, often far exceeding available GPU resources and context window sizes. Our key observation is that agents need only send what the receiving agent requires for its local task -- which we call receiver-conditioned communication. The receiver agent passes the sender a small description of its information needs, which serves to filter and compress the sender agent's KV cache. CacheBack is a simple, robust, training-free instance of receiver conditioning based on the sender's attention weights. On FanOutQA, CacheBack with Qwen 3 removes 75% of the state the agent would otherwise receive, improving accuracy by 14.7 percentage points and reducing median task-completion latency by 3.2x relative to text communication. We show comparable improvements across model families that span dense Transformers, Mamba-attention hybrids, and sliding-window attention.

cs.AI↗
arXiv2026

Evolving Safety Landscape of Multi-modal Large Language Models: A Survey of Emerging Threats and Safeguards

Multi-modal large language models (MLLMs) integrate heterogeneous modalities through modality alignment and fusion, enabling stronger understanding and reasoning. However, this architectural shift reshapes the safety landscape of machine learning. Increased model complexity and cross-modal interactions give rise to novel threats, including compromised modality integration, modality misalignment, and fused safety risks, reflecting shifts in threat modeling beyond uni-modal assumptions. These shifts, in turn, impose new constraints on safety solutions not captured by existing frameworks rooted in uni-modal learning. Motivated by these challenges, this survey provides a systematic analysis of the evolving safety landscape of MLLMs. We first propose a multimodal grounded taxonomy of safety threats and analyze shifts in threat models, covering adversarial attacks, data poisoning, jailbreaks, and hallucinations. We then summarize updated safety assumptions and organize recent advances in MLLM safety strategies accordingly. Finally, we discuss open challenges and future directions to inform the development of more principled and scalable safety mechanisms for multimodal systems.

cs.LG↗
arXiv2026

Agentic Data Environments

Autonomous agents promise substantial gains in speed, scale, and labor efficiency, but their failures can impose abrupt and often irreversible costs. The central challenge for agentic automation is therefore to increase the benefits of automation while bounding the consequences of failure. While databases remain central to modern computing, agents operate over a broader data environment spanning files, APIs, applications, and system state. In this talk, I will outline early work on Agentic Data Environments -- the execution substrate in which agents operate -- that both amplify agent capabilities and enforce safety guarantees. This perspective reframes data systems from passive stores of state into active substrates for safe, reliable execution.

cs.AI↗
arXiv2026

LakeQA: An Exploratory QA Benchmark over a Million-Scale Data Lake

Recent large language models (LLMs) have shown rapid progress in reading-based question answering (QA), where evidence is explicitly provided or can be trivially retrieved. In contrast, real-world questions are often not paired with accurate evidence documents. The useful evidence resides in massive data lakes, making search a prerequisite for answering. However, there is a lack of comprehensive benchmarks that require both searching and reasoning over large data lakes. To this end, we introduce LakeQA, a comprehensive benchmark for search-centric question answering over data lakes that jointly emphasizes searching and reasoning capabilities. LakeQA is built on a heterogeneous collection of approximately 9.5 TB of text resources from Wikipedia and open-source government data, spanning structured and unstructured data. To ensure task quality, each sample is annotated by at least one Ph.D.-level expert. Each task requires long-horizon multi-hop reasoning with implicit intermediate steps: agents need to discover the correct documents and then compose evidence across sources to produce the answer. Experimental results on seven frontier LLMs demonstrate that LakeQA is challenging. For instance, GPT-5.2 achieves only an exact-match score of 18.37% on LakeQA. Overall, LakeQA provides a realistic testbed for developing LLM agents that can both find and analyze data in modern data lakes.

cs.CL↗
arXiv2026

ReMoE: Boosting Expert Reuse through Router Fine-Tuning in Memory-Constrained MoE LLM Inference

Fine-grained Mixture-of-Experts (MoE) models sparsely activate only a subset of experts per token, reducing activated computation while maintaining high model capacity. However, in memory-constrained inference scenarios, only a small set of experts can be cached. Experts not in the cache must be fetched from slow external storage (e.g., UFS), leading to frequent evictions and substantial I/O overhead. We propose ReMoE, a router fine-tuning framework designed to boost token-wise expert reuse. ReMoE biases the router toward recently selected experts, producing temporally stable routing that better matches cache locality constraints. By increasing short-horizon expert reuse, ReMoE reduces expert fetches from storage without adding inference-time computation. Experiments on DeepSeek and Qwen models show that ReMoE improves expert reuse by 26% while maintaining downstream task performance. Real-system evaluations further confirm these benefits, improving output throughput by 8.4% under vLLM GPU-CPU expert offloading and reducing TPOT by 43.6-49.8% under llama.cpp on Jetson Orin NX, corresponding to a 1.77-1.99$\times$ decode speedup across diverse workloads. Checkpoints and usage instructions are available at https://github.com/BUAA-OSCAR/ReMoE.

cs.LG↗
arXiv2026

Determinism of Randomness: Prompt-Residual Seed Shaping for Diffusion Generation

Diffusion models start generation from an isotropic Gaussian latent, yet changing only the random seed can lead to large differences in prompt faithfulness, composition, and visual quality. We study this seed sensitivity through the semantic map from initial noise to generated meaning. Although the sampling flow is locally invertible, the subsequent semantic projection is many-to-one, inducing a degenerate pullback semi-metric on the latent space: most local directions are nearly semantic-invariant, while semantic-sensitive variation is concentrated in a much smaller horizontal subspace. This provides an explanatory geometric view of the seed lottery. Motivated by this view, we introduce a training-free prompt-residual seed-shaping procedure. Rather than claiming to recover the exact horizontal space, the method uses a single high-noise cold-start prompt residual as a model-coupled proxy, injects only its tangential component, and retracts the seed to the original Gaussian radius shell. This keeps the initialization prior-compatible while adding only one conditional/unconditional probe before standard sampling. Across multiple generation benchmarks, the method improves alignment and quality metrics over standard sampling, supporting both the practical value of the proxy and the explanatory relevance of semantic anisotropy.

cs.CV↗
arXiv2026

Domain-Adaptive Model Merging Across Disconnected Modes

Learning across domains is challenging when data cannot be centralized due to privacy or heterogeneity, which limits the ability to train a single comprehensive model. Model merging provides an appealing alternative by consolidating knowledge from multiple specialized models into one, avoiding data sharing and reducing retraining cost. In this work, we present DMM, a data-free model merging framework designed to handle highly divergent models. DMM proceeds in three steps. First, domain-specific models are trained independently. Second, models with high similarity are merged using standard techniques to ensure stability. Third, we synthesize pseudo-data from normalization statistics and distill knowledge from divergent models into the merged model through a lightweight refinement guided by these samples. This approach preserves rare but critical knowledge while maintaining stability. Extensive experiments on unimodal and multimodal benchmarks show that DMM achieves state-of-the-art performance over existing merging methods.

cs.DC↗
arXiv2026

Efficient PRM Training Data Synthesis via Formal Verification

Process Reward Models (PRMs) have emerged as a promising approach for improving LLM reasoning capabilities by providing process supervision over reasoning traces. However, existing approaches for constructing PRM training data remain costly and noisy, as they typically rely on human annotation or sampling-based labeling methods that require repeated LLM calls. In this work, we propose FoVer, a framework that synthesizes PRM training data from formal reasoning tasks by annotating step-level error labels using formal verification tools such as Z3 and Isabelle. By leveraging formal verification, FoVer enables efficient and accurate PRM data construction without requiring human annotation or additional LLM calls. Using FoVer, we create PRM training data from formal logic and theorem proving tasks. Experiments on 12 reasoning benchmarks show that fine-tuning on our training data improves PRMs not only on math and logic reasoning tasks, which are informal variants of the training tasks, but also on NLI and BBH benchmarks, which differ substantially from the tasks used to construct the training data. These results demonstrate the practical effectiveness of FoVer, showing that PRM training data created using formal verification improves PRMs on informal reasoning tasks written in natural language. The datasets, models, and code are provided at https://github.com/psunlpgroup/FoVer.

cs.CL↗
arXiv2026

Bridging the Know-Act Gap via Task-Level Autoregressive Reasoning

LLMs often generate seemingly valid answers to flawed or ill-posed inputs. This is not due to missing knowledge: under discriminative prompting, the same models can mostly identify such issues, yet fail to reflect this in standard generative responses. This reveals a fundamental know-act gap between discriminative recognition and generative behavior. Prior work largely characterizes this issue in narrow settings, such as math word problems or question answering, with limited focus on how to integrate these two modes. In this work, we present a comprehensive analysis using FaultyScience, a newly constructed large-scale, cross-disciplinary benchmark of faulty scientific questions. We show that the gap is pervasive and stems from token-level autoregression, which entangles task selection (validate vs. answer) with content generation, preventing discriminative knowledge from being utilized. To address this, we propose DeIllusionLLM, a task-level autoregressive framework that explicitly models this decision. Through self-distillation, the model unifies discriminative judgment and generative reasoning within a single backbone. Empirically, DeIllusionLLM substantially reduces answer-despite-error failures under natural prompting while maintaining general reasoning performance, demonstrating that self-distillation is an effective and scalable solution for bridging the discriminative-generative know-act gap

cs.AI↗
arXiv2026

When Reasoning Meets Compression: Understanding the Effects of LLMs Compression on Large Reasoning Models

Compression methods, including quantization, distillation, and pruning, improve the computational efficiency of large reasoning models (LRMs). However, existing studies either fail to sufficiently compare all three compression methods on LRMs or lack in-depth interpretation analysis. In this paper, we investigate how the reasoning capabilities of LRMs are compromised during compression, through performance benchmarking and mechanistic interpretation. To uncover the effects of compression on reasoning performance, we benchmark quantized, distilled, and pruned DeepSeek-R1 models on four reasoning datasets (AIME 2024, FOLIO, Temporal Sequences, and MuSiQue). To precisely locate compression effects on model weights, we adapt difference of means and attribution patching techniques, focusing on the activation of every linear component in compressed LRMs, to interpret fine-grained causal relationships between weights and various reasoning capabilities. This fine-grained interpretation addresses a fundamental question of compression: which weights are the most important for reasoning? Overall, we find dynamically quantized 2.51-bit R1 reaches close-to-R1 performance. With empirical verification, we present three main findings that generalize across both R1 and non-R1 LRMs: (1) Weight count has a greater impact on LRMs' knowledge memorization than reasoning, highlighting the risks of pruning and distillation; (2) The MLP up projection in the final layer of distilled LRMs is one of the most important components, offering a new perspective on locating critical weights - a fundamental problem in model compression; and (3) Current quantization methods overly compress the final-layer modules and MLP gate projections, so protecting just 2% of all weights that are excessively compressed can raise average accuracy by 6.57%, greatly surpassing the state-of-the-art.

cs.LG↗
arXiv2026

QuantLRM: Quantization of Large Reasoning Models via Fine-Tuning Signals

Weight-only quantization is important for compressing Large Language Models (LLMs). Inspired by the spirit of classical magnitude pruning, we study whether the magnitude of weight updates during reasoning-incentivized fine-tuning can provide valuable signals for quantizing Large Reasoning Models (LRMs). We hypothesize that the smallest and largest weight updates during fine-tuning are more important than those of intermediate magnitude, a phenomenon we term "protecting both ends". Upon hypothesis validation, we introduce QuantLRM, which stands for weight quantization of LRMs via fine-tuning signals. We fit simple restricted quadratic functions on weight updates to protect both ends. By multiplying the average quadratic values with the count of zero weight updates of channels, we compute channel importance that is more effective than using activation or second-order information. We run QuantLRM to quantize various fine-tuned models (including supervised, direct preference optimization, and reinforcement learning fine-tuning) over four reasoning benchmarks (AIME-120, FOLIO, temporal sequences, and GPQA-Diamond) and empirically find that QuantLRM delivers a consistent improvement for LRMs quantization, with an average improvement of 6.55% on a reinforcement learning fine-tuned model. Also supporting non-fine-tuned LRMs, QuantLRM gathers effective signals via pseudo-fine-tuning, which greatly enhances its applicability.

cs.LG↗
arXiv2026

An Improved Reversible Data Hiding Algorithm Based on Reconstructed Mapping for PVO-k

Reversible Data Hiding (RDH) is a practical and efficient technique for information encryption. Among its methods, the Pixel-Value Ordering (PVO) algorithm and its variants primarily modify prediction errors to embed information. However, both the classic PVO and its improved versions, such as IPVO and PVO-k, share a common limitation: their maximum data embedding capacity for a given grayscale image is relatively low. This poses a challenge when large amounts of data need to be embedded into an image. In response to these issues, this paper proposes an improved design targeting the PVO-k algorithm. We have reconstructed the mapping scheme of the PVO-k algorithm to maximize the number of pixels that can embed encrypted information. Experimental validations show that our proposed scheme significantly surpasses previous algorithms in terms of the maximum data embedding capacity. For instance, when embedding information into a grayscale image of an airplane, our method's capacity exceeds that of PVO-k by 11,207 bits, PVO by 8,004 bits, and IPVO by 4,562 bits. The results demonstrate that our algorithm holds substantial advantages over existing methods and introduces innovative mapping ideas, laying a foundation for future research in reversible data hiding in images.

cs.CR↗
arXiv2025

Chain-of-Scrutiny: Detecting Backdoor Attacks for Large Language Models

Large Language Models (LLMs), especially those accessed via APIs, have demonstrated impressive capabilities across various domains. However, users without technical expertise often turn to (untrustworthy) third-party services, such as prompt engineering, to enhance their LLM experience, creating vulnerabilities to adversarial threats like backdoor attacks. Backdoor-compromised LLMs generate malicious outputs to users when inputs contain specific "triggers" set by attackers. Traditional defense strategies, originally designed for small-scale models, are impractical for API-accessible LLMs due to limited model access, high computational costs, and data requirements. To address these limitations, we propose Chain-of-Scrutiny (CoS) which leverages LLMs' unique reasoning abilities to mitigate backdoor attacks. It guides the LLM to generate reasoning steps for a given input and scrutinizes for consistency with the final output -- any inconsistencies indicating a potential attack. It is well-suited for the popular API-only LLM deployments, enabling detection at minimal cost and with little data. User-friendly and driven by natural language, it allows non-experts to perform the defense independently while maintaining transparency. We validate the effectiveness of CoS through extensive experiments on various tasks and LLMs, with results showing greater benefits for more powerful LLMs.

cs.CR↗
arXiv2025

VisOnlyQA: Large Vision Language Models Still Struggle with Visual Perception of Geometric Information

Large Vision Language Models (LVLMs) have achieved remarkable performance in various vision-language tasks. However, it is still unclear how accurately LVLMs can perceive visual information in images. In particular, the capability of LVLMs to perceive geometric information, such as shape, angle, and size, remains insufficiently analyzed, although the perception of these properties is crucial for tasks that require a detailed visual understanding. In this work, we introduce VisOnlyQA, a dataset for evaluating the geometric perception of LVLMs, and reveal that LVLMs often cannot accurately perceive basic geometric information in images, while human performance is nearly perfect. VisOnlyQA consists of 12 tasks that directly ask about geometric information in geometric shapes, charts, chemical structures, and 3D shapes. Our experiments highlight the following findings: (i) State-of-the-art LVLMs struggle with basic geometric perception. 23 LVLMs we evaluate, including GPT-4o and Gemini 2.5 Pro, work poorly on VisOnlyQA. (ii) Additional training data does not resolve this issue. Fine-tuning on the training set of VisOnlyQA is not always effective, even for in-distribution tasks. (iii) LLM may be the bottleneck. LVLMs using stronger LLMs exhibit better geometric perception on VisOnlyQA, while it does not require complex reasoning, suggesting that the way LVLMs process information from visual encoders is a bottleneck. The datasets, code, and model responses are provided at https://github.com/psunlpgroup/VisOnlyQA.

cs.CL↗
arXiv2025

Break Stylistic Sophon: Are We Really Meant to Confine the Imagination in Style Transfer?

In this pioneering study, we introduce StyleWallfacer, a groundbreaking unified training and inference framework, which not only addresses various issues encountered in the style transfer process of traditional methods but also unifies the framework for different tasks. This framework is designed to revolutionize the field by enabling artist level style transfer and text driven stylization. First, we propose a semantic-based style injection method that uses BLIP to generate text descriptions strictly aligned with the semantics of the style image in CLIP space. By leveraging a large language model to remove style-related descriptions from these descriptions, we create a semantic gap. This gap is then used to fine-tune the model, enabling efficient and drift-free injection of style knowledge. Second, we propose a data augmentation strategy based on human feedback, incorporating high-quality samples generated early in the fine-tuning process into the training set to facilitate progressive learning and significantly reduce its overfitting. Finally, we design a training-free triple diffusion process using the fine-tuned model, which manipulates the features of self-attention layers in a manner similar to the cross-attention mechanism. Specifically, in the generation process, the key and value of the content-related process are replaced with those of the style-related process to inject style while maintaining text control over the model. We also introduce query preservation to mitigate disruptions to the original content. Under such a design, we have achieved high-quality image-driven style transfer and text-driven stylization, delivering artist-level style transfer results while preserving the original image content. Moreover, we achieve image color editing during the style transfer process for the first time.

cs.CV↗