Searcharxiv⌕ Search

arXiv subjects

Nanyun Peng

Publications and source records attributed to Nanyun Peng.

At least 19 recordsLinked to original sources
arXiv2026

TwinICL: Diagnosing Multimodal In-Context Learning through Paired Counterfactuals

In-context learning (ICL) enables models to infer tasks from demonstrations, but existing benchmarks generally lack matched text and image versions needed to compare ICL performance across modalities. We introduce TwinICL, a procedurally generated benchmark providing such pairs for controlled comparison. Across six open-weight models and 38 tasks, multimodal ICL consistently underperforms text-only ICL, with gaps varying by task family. To test whether this gap can be recovered, we target visual access, task framing, and reasoning through three interventions. Their combination recovers strong multimodal ICL performance on a diagnostic subset, despite limited or inconsistent individual effects. To distinguish difficulties in executing tasks from those in inferring them, we evaluate models with explicit task instructions, revealing a modality gap even when the task is known. We then examine how adding demonstration inputs and outputs reshapes this gap, highlighting demonstrations' dual role as additional context to process and evidence about the task. The dataset is available at https://github.com/lab-flair/TwinICL.

cs.CV↗
arXiv2026

Does Reasoning Improve Psychological Depth in Large Language Models? It Depends on Who's Judging

LLM-as-a-Judge evaluators are increasingly used to score open-ended generation, yet a judge's correlation with human ratings on its development set may not guarantee valid measurement when outputs are closely matched and human preferences are subjective. We study this failure mode through psychological depth in short stories. Seven human readers and an LLM-judge ensemble selected on the original scalar Psychological Depth Scale dataset ($ρ= 0.646$) evaluated 60 blinded, prompt-matched story pairs from GPT-5 vs.\ GPT-4o and DeepSeek-R1 vs.\ DeepSeek-V3. Human preferences showed no universal reasoning advantage: GPT-5 was modestly preferred over GPT-4o (60.0--62.9\%), whereas DeepSeek-R1 trailed V3 (42.9\%), and inter-reader agreement was near chance (Krippendorff's $α= 0.070$), with within-reader consistency and recurring weighting patterns suggesting structured heterogeneity rather than random responding. The judge, by contrast, favored reasoning outputs in 89.0\% of dimension-level comparisons and 59 of 60 pairs on aggregate PDS, uniformly across all five evaluator configurations, and its scores were associated with surface features such as sentence length and lexical diversity. These results suggest that development-set performance is insufficient evidence for deployment validity on a shifted distribution, and that point-estimate judges can obscure the heterogeneity in subjective human evaluation.

cs.LG↗
arXiv2026

ESTS at WMT26: Routing-Informed Expert Pruning for Model Compression

We describe six submissions under the team name ESTS to the unconstrained WMT26 Model Compression Shared Task for English--Simplified Chinese and English--Egyptian Arabic. We submit three compression operating points per translation direction, all derived from GPT-OSS-20B. We use task-specific routing mass to rank experts and cross-lingual routing divergence to allocate retained capacity across layers, then physically remove low-importance experts. The resulting specialists are recovery-tuned on GPT-5.1-generated synthetic translation data and further compressed by applying MXFP4 quantization to the retained expert projection weights. We additionally implement a robust inference system for the instruction-conditioned WMT26 setting, including category inference, output validation, retries, segmented fallback, and source-owned JSON reconstruction. Across our six submissions, parameter counts range from 4.186B to 7.770B and packed artifact sizes from 4.55 to 6.33~GiB. Internal xCOMET-XL evaluation using GPT-5.1 pseudo-references provides an internal comparison across the submitted compression operating points.

cs.CL↗
arXiv2026

MMGR: Multi-Modal Generative Reasoning Benchmark and Evaluation

Modern multimodal generative models can synthesize visually compelling images and videos, but it remains unclear whether this visual fluency reflects genuine reasoning: when prompted to generate a solution, can a model preserve the physical, logical, spatial, and temporal constraints a task requires, or does it merely produce plausible-looking media? To answer this question, we introduce MMGR (Multi-Modal Generative Reasoning Benchmark and Evaluation), a benchmark for evaluating generative reasoning across video, image, and language-based systems. MMGR covers 10 tasks from three domains (Abstract Reasoning, Embodied Navigation, and Physical Commonsense) and probes five reasoning abilities: Physical, Logical, 2D Spatial, 3D Spatial, and Temporal. Its evaluation emphasizes answer-verifiable tasks and, for video generation, process-aware chain-of-frame reasoning, where intermediate frames must form valid steps toward the target outcome rather than visually smooth but incorrect transitions. Evaluating state-of-the-art video generators, image generators, and LLM/VLM baselines reveals a sharp gap between visual quality and reasoning correctness: video models perform best on Physical Commonsense, but remain weak on symbolic tasks such as Sudoku, ARC, and Math, and brittle in cross-view embodied navigation. Image generators often outperform video generators on embodied navigation despite lacking temporal outputs, showing that longer visual generation does not automatically yield stronger reasoning. MMGR reframes evaluation of multimodal generation from whether outputs look realistic to whether they solve the underlying reasoning problem.

cs.CL↗
arXiv2026

NormViz: A Benchmark and Framework for Grounding Multimodal Reasoning in Global Cultures

AI systems are used worldwide, but they struggle to serve the needs of culturally diverse populations. Prior work on cultural understanding evaluates AI systems on text-only settings or on visual artifact recognition (e.g. foods, clothing). The ability to reason about visually observable behaviors through local social norms, which we call visual norm understanding, remains unexamined. We introduce NormViz-Bench, a high quality, human-validated benchmark of 3,268 contrastive image pairs (6,536 images) spanning 16 countries. Each pair varies only in the culturally relevant behavior (e.g., objects, attributes, spatial relations, and actions) that alters how each image is interpreted. Each image is labeled as conforming to, violating, or irrelevant to local social norms, and pair-level evaluation requires both images to be correctly classified, thereby preventing reliance on superficial visual shortcuts. Even the strongest VLMs, Gemini 3.0 Flash and Qwen2.5 VL 7B, succeed on only 26.6% and 21.6% of pairs, struggling most with identifying violating and culturally benign visual behaviors. Towards bridging this, we introduce NormViz-Train, a training dataset of 64k images paired with explanations. Though absolute performance remains low (<30%), finetuning on NormViz-Train improves pair accuracy relatively by up to 125% and 36% Qwen3-VL 4B and 8B respectively, showing a path forward to teach models to connect visual perception to cultural significance. Together, NormViz-Bench and NormViz-Train establish visual norm understanding as a challenging and consequential frontier for multimodal AI.

cs.AI↗
arXiv2026

On Asymmetric Optimization of Reasoning and Perception in Vision-Language Model Post-Training

Post-training has greatly improved reasoning in frontier vision-language models, yet its gains for perception remain comparatively limited, creating a bottleneck for end-to-end visual reasoning. To investigate this gap, we introduce a controlled diagnostic framework with two synthetic tasks that disentangle perception from reasoning. Our analysis reveals a consistent perception-reasoning asymmetry: post-training improves reasoning more substantially than perception, though the underlying mechanism differs across training paradigms. For supervised fine-tuning (SFT), this asymmetry stems from token imbalance, with perception occupying a smaller fraction of tokens in chain-of-thought supervision. Reweighting the loss boosts end-to-end performance by up to 18.2 points. For reinforcement learning (RL), the asymmetry instead arises from reward coupling, as outcome rewards correlate more strongly with reasoning than perception. Adding a perception-aware reward improves end-to-end accuracy by up to 6.0 points; when ground-truth perception rewards are unavailable, a reliable surrogate provides useful signal, yielding gains of 2.2 points. Beyond the controlled setting, these strategies also improve real-world visual reasoning, with gains of up to 3.3 points across three benchmarks. Overall, we diagnose the causes of asymmetric optimization and provide actionable guidance that benefits both synthetic and realistic settings.

cs.CL↗
arXiv2026

ScalePRM: Training Process Reward Models by Scaling Verification Compute Without Ground Truth

Training process reward models (PRMs) requires step-level correctness labels, obtained either through expensive human annotation or by relying on ground-truth answers, limiting the ability to scale process-level supervision. We propose ScalePRM, which scales verification compute as an alternative: given a problem and a candidate solution, we generate multiple independent verifications of each reasoning step and aggregate their judgments to produce synthetic step-level labels without ground truth. We explore two representative inference-time scaling strategies, parallel scaling through self-consistency and sequential scaling through meta-critique, and train generative PRMs on the resulting synthetic data. On ProcessBench, a benchmark for identifying erroneous steps in mathematical reasoning, PRMs trained on step-level self-consistency data achieve 67.5 F1, surpassing reference-guided training with ground-truth access (66.4 F1) and GPT-4o as a critic (61.9 F1). When deployed as reward signals in RL training with Qwen2.5-Math-7B, our best PRM achieves 47.4% average accuracy across six mathematical reasoning benchmarks, outperforming ground-truth-based RLVR (43.9%). We also identify and address reward exploitation patterns unique to generative PRM-based RL. Our results demonstrate that scaling verification compute is a viable alternative to ground-truth supervision for training process reward models.

cs.LG↗
arXiv2026

Extracting Small Translation Specialists from LLMs by Aggressively Pruning Experts

Modern large language models (LLMs) achieve state-of-the-art machine translation performance, but they do so as broad generalists largely trained for many tasks and capabilities unrelated to translation. Thus, they are heavily overparameterized for this task, resulting in excessive memory and compute requirements. In this paper, we present a method for aggressively pruning experts from modern mixture-of-experts LLMs while incurring negligible degradation in translation quality. Our approach exploits expert specialization and the separability of multilingual capabilities in LLMs to identify experts irrelevant to translation. And because of the modular nature of MoEs, these can be easily pruned without any training. Without retraining, we are able to prune half of all experts with negligible degradation and 70% with only minor losses. With a very short SFT, we prune 75% of experts while recovering baseline performance, and in some settings remove nearly 90% while maintaining reasonable translation quality. Overall, our results show that translation requires only a fraction of the LLM, enabling substantial compression of the MoE blocks that contain over 90% of parameters.

cs.CL↗
arXiv2026

PaperBanana-Interact: Scientific Diagram Refinement with Multi-Turn Human Feedback

Recent efforts have aimed to automate scientific diagram generation from paper content (Lin et al., 2026; Zhu et al., 2026a). However, fully satisfying an author's visual and communicative preferences in a single turn is challenging: in our formative user study (N = 14), all participants requested further revisions after viewing an initial draft, and 86% of them rated the refined diagrams as more satisfactory. Despite the clear demand, the multi-turn workflow remains largely underexplored. To bridge this gap, we present MTPaperBananaBench, a benchmark for multi-turn diagram generation containing 292 images annotated with 3,518 user requirements. To reduce expensive human studies and enable scalable benchmarking, we construct a user simulator that, at each turn, identifies unsatisfied requirements and converts k of them into natural language feedback. Evaluating both requirement satisfaction and overall diagram quality reveals two key failure modes shared across baseline multiturn systems: (1) quality drift, where diagram quality progressively declines over turns, and (2) forgetting, where previously implemented features are lost in subsequent turns. To address these issues, we introduce PaperBanana-Interact, a multi-agent system that refines diagrams via an internal critique-and-refine loop. PaperBanana-Interact consistently improves rather than degrades diagram quality across turns, outperforming baselines by 11.9-18.6 points in quality score and reducing forgetting by 3.7-6.2 points.

cs.CL↗
arXiv2026

Belief Cascades Drive Persuasion in LLM Agent Networks

Multi-agent LLM systems increasingly debate answers, coordinate research, simulate users, and mediate information flows, making agent-to-agent persuasion a basic but undermeasured capability. We introduce a controlled testbed for studying how goal-directed persuaders shift elicited stances in networks of LLM agents grounded in real-world ego-network topologies. Across four LLM backbones, five graphs, and 55 policy statements, we find that persuasion dynamics depend on the interaction between topology, competition, topic, and model prior. Additionally, we show that direct exposure reliably predicts next-round stance change in competing runs, and peer relays carry smaller but measurable influence, showing that agents not assigned to persuade can still transmit persuasive force. Finally, analyzing post text alone misses important movement: planned strategies are only partly realized in executed messages, action choices can diverge from message content, and persuadees rarely state the stance shifts detected by probes. These results argue for evaluating multi-agent persuasion as a trajectory- and exposure-level process, using belief probes, exposure provenance, and action logs to identify who influenced whom and whether visible language reflects underlying stance movement.

cs.CL↗
arXiv2026

TemMed-Bench: Evaluating Temporal Medical Image Reasoning in Vision-Language Models

Existing medical reasoning benchmarks for vision-language models primarily focus on analyzing a patient's condition based on an image from a single visit. However, this setting deviates significantly from real-world clinical practice, where doctors typically refer to a patient's historical conditions to provide a comprehensive assessment by tracking their changes over time. In this paper, we introduce TEMMED-BENCH, a multi-task benchmark designed for analyzing changes in patients' conditions between different clinical visits, which challenges large vision-language models (LVLMs) to reason over temporal medical images. TEMMED-BENCH consists of a test set comprising three tasks - visual question-answering (VQA), report generation, and image-pair selection - and a supplementary knowledge corpus of over 17,000 instances. With TEMMED-BENCH, we conduct an evaluation of twelve LVLMs, comprising six proprietary and six open-source models. Our results show that most LVLMs lack the ability to analyze patients' condition changes over temporal medical images, and a large proportion perform only at a random-guessing level in the closed-book setting. To enhance the tracking of condition changes, we explore augmenting the input with both retrieved visual and textual modalities in the medical domain. We also show that multi-modal retrieval augmentation yields notably higher performance gains than no retrieval and textual retrieval alone across most models on our benchmark, with the VQA task showing an average improvement of 2.59%. Overall, we compose a benchmark grounded on real-world clinical practice, and it reveals LVLMs' limitations in temporal medical image reasoning, as well as highlighting the use of multi-modal retrieval augmentation as a potentially promising direction worth exploring to address this challenge.

cs.CV↗
arXiv2026

FronTalk: Benchmarking Front-End Development as Conversational Code Generation with Multi-Modal Feedback

We present FronTalk, a benchmark for front-end code generation that pioneers the study of a unique interaction dynamic: conversational code generation with multi-modal feedback. In front-end development, visual artifacts such as sketches, mockups and annotated creenshots are essential for conveying design intent, yet their role in multi-turn code generation remains largely unexplored. To address this gap, we focus on the front-end development task and curate FronTalk, a collection of 100 multi-turn dialogues derived from real-world websites across diverse domains such as news, finance, and art. Each turn features both a textual instruction and an equivalent visual instruction, each representing the same user intent. To comprehensively evaluate model performance, we propose a novel agent-based evaluation framework leveraging a web agent to simulate users and explore the website, and thus measuring both functional correctness and user experience. Evaluation of 20 models reveals two key challenges that are under-explored systematically in the literature: (1) a significant forgetting issue where models overwrite previously implemented features, resulting in task failures, and (2) a persistent challenge in interpreting visual feedback, especially for open-source vision-language models (VLMs). We propose a strong baseline to tackle the forgetting issue with AceCoder, a method that critiques the implementation of every past instruction using an autonomous web agent. This approach significantly reduces forgetting to nearly zero and improves the performance by up to 9.3% (56.0% to 65.3%). Overall, we aim to provide a solid foundation for future research in front-end development and the general interaction dynamics of multi-turn, multi-modal code generation. Code and data are released at https://github.com/shirley-wu/frontalk

cs.CL↗
arXiv2026

LLM Self-Correction with DeCRIM: Decompose, Critique, and Refine for Enhanced Following of Instructions with Multiple Constraints

Instruction following is a key capability for LLMs. However, recent studies have shown that LLMs often struggle with instructions containing multiple constraints (e.g. a request to create a social media post "in a funny tone" with "no hashtag"). Despite this, most evaluations focus solely on synthetic data. To address this, we introduce RealInstruct, the first benchmark designed to evaluate LLMs' ability to follow real-world multi-constrained instructions by leveraging queries real users asked AI assistants. We also investigate model-based evaluation as a cost-effective alternative to human annotation for this task. Our findings reveal that even the proprietary GPT-4 model fails to meet at least one constraint on over 21% of instructions, highlighting the limitations of state-of-the-art models. To address the performance gap between open-source and proprietary models, we propose the Decompose, Critique and Refine (DeCRIM) self-correction pipeline, which enhances LLMs' ability to follow constraints. DeCRIM works by decomposing the original instruction into a list of constraints and using a Critic model to decide when and where the LLM's response needs refinement. Our results show that DeCRIM improves Mistral's performance by 7.3% on RealInstruct and 8.0% on IFEval even with weak feedback. Moreover, we demonstrate that with strong feedback, open-source LLMs with DeCRIM can outperform GPT-4 on both benchmarks.

cs.CL↗
arXiv2026

Decoupling Task-Solving and Output Formatting in LLM Generation

Large language models (LLMs) are increasingly adept at solving complex problems, such as mathematical reasoning and automatic evaluation. However, performance often degrades when prompts intertwine task instructions with rigid formatting requirements. This entanglement creates competing goals for the model, hindering its reasoning capabilities. To address this, we introduce Deco-G, a decoding framework that explicitly decouples format adherence from problem solving. Deco-G delegates format adherence to a separate Format Estimation Module (FEM), which performs probabilistic lookahead to estimate future format compliance rate and reweighs token probabilities, allowing the LLM to focus solely on task resolution. To make this approach both practical and efficient, we introduce three key innovations: instruction-aware distillation, a flexible trie-building algorithm, and HMM state pruning. Experiments across mathematical reasoning, event argument extraction, and LLM-as-a-judge demonstrate that Deco-G constantly gains over prompting or structured generation baselines, with guaranteed format compliance. We release our code at https://github.com/haikangdeng/deco-g.

cs.CL↗
arXiv2026

From Solvers to Research: Large Language Model-Driven Formal Mathematics at the Research Frontier

Recent developments in AI for Mathematics (AI4Math), especially Large Language Model (LLM)-driven theorem provers, has achieved remarkable success in formal proof generation for well-defined mathematical problems through Interactive Theorem Proving (ITP) languages. However, current systems remain fundamentally limited in tackling frontier research mathematics, such as discovering new theorems or resolving open conjectures, which are often open-ended, under-specified, and involve multiple layers of abstraction. We argue that the next leap in AI4Math systems requires a decisive shift from predefined problem-solvers to research agents that can address frontier mathematical challenges with rigorous formal mathematical reasoning. In this position paper, we provide a systematic review of the field, covering datasets, auto-formalization, and proof synthesis. More importantly, we identify core limitations of existing systems in serving as mathematical research agents, examining issues across datasets, relational structure, mathematical exploration, tool ecosystem, and human-AI collaboration, outlining a strategic road-map for the future of AI4Math.

cs.CL↗
arXiv2026

LEAP: Supercharging LLMs for Formal Mathematics with Agentic Frameworks

Large Language Models (LLMs) exhibit strong informal mathematical reasoning but struggle to generate mechanically verifiable proofs in formal languages like Lean. We present LEAP, an agentic framework that enables general-purpose foundation models to achieve state-of-the-art performance on automated formal theorem proving. LEAP leverages foundation model capabilities, such as informal reasoning, instruction following, and iterative self-refinement. By decomposing complex problems into smaller units, the system bridges formal proof construction with informal blueprints through continuous interaction with the Lean compiler. To provide a rigorous evaluation beyond increasingly saturated benchmarks, we introduce Lean-IMO-Bench, a benchmark of IMO-style problems formalized in Lean, with short statements yet highly non-routine and multi-step proofs across a wide range of difficulty levels. Empirically, on the latest 2025 Putnam Competition, an annual mathematics competition for undergraduate students in North America, LEAP solves all 12 problems, matching recent breakthroughs by frontier formal mathematical models. On Lean-IMO-Bench, LEAP boosts the one-shot formal solve rate of general-purpose LLMs from below 10% to 70%, notably surpassing the 48% benchmark set by a specialized, gold-medal-caliber IMO system. Furthermore, we demonstrate LEAP's research-level utility by autonomously formalizing complex proofs for open combinatorial challenges, including a verified proof for a key subproblem in Knuth's Hamiltonian decomposition of even-order Cayley graphs.

cs.AI↗
arXiv2026

Seeing is Believing? Evaluating Vision-Language Model Susceptibility in Agent-to-Agent Multimodal Persuasion

As autonomous agents increasingly interact, they inevitably attempt to influence one another. While prior work in text-only settings has explored the dynamics of Agent-to-Agent (A2A) persuasion, the rise of Vision-Language Models (VLMs) introduces a more complex challenge: multimodal content conveys richer information while integrating subtle, hard-to-detect persuasive cues. To study this vulnerability, we present MMPersuade, a unified framework and dataset for A2A multimodal persuasion. We model interactions between a persuader agent, which leverages images and psychological strategies, and a persuadee VLM. Our benchmark spans commercial, subjective and behavioral, and adversarial contexts, and evaluates persuasion via function-calling that capture behavioral shifts beyond verbal responses. Experiments on six VLMs reveal three findings: (1) multimodal inputs consistently outperform text-only persuasion, with raw visual signals uniquely increasing susceptibility in adversarial settings by bypassing text-activated safety defenses; (2) persuadee vulnerability is highly domain- and format-dependent, with realistic and community-style formats driving susceptibility in commercial settings while different formats dominate in adversarial ones; and (3) psychological strategy efficacy varies with context and model architecture, as more capable models resist benign persuasion yet become more susceptible under adversarial multimodal inputs. Our framework provides a foundation for building more robust and aligned VLMs in multi-agent environments.

cs.CL↗
arXiv2026

MM-PoisonRAG: Disrupting Multimodal RAG with Local and Global Poisoning Attacks

Retrieval-augmented generation (RAG) has become a common practice in multimodal large language models (MLLM) to enhance factual grounding and reduce hallucination. Yet, its reliance on retrieval exposes MLLMs to knowledge poisoning attacks, in which adversaries deliberately inject malicious multimodal content into external knowledge bases to steer models toward generating incorrect or even harmful responses. We present MM-PoisonRAG, a framework to systematically study the vulnerability of multimodal RAG under knowledge poisoning. Specifically, we design two novel attack strategies: Localized Poisoning Attack (LPA), which implants targeted, query-specific multimodal misinformation to manipulate outputs toward attacker-controlled responses, and Globalized Poisoning Attack (GPA), which uses a single, untargeted adversarial injection to broadly corrupt reasoning and collapse generation quality across all queries. Extensive experiments on diverse tasks, multimodal RAG components, and attacker access levels reveal severe vulnerabilities: LPA achieves up to 56% attack success rate even under restricted access, and transfers effectively across four different retrievers without re-optimizing the adversaries. GPA completely disrupts model generation to 0% accuracy with just one poisoned content. Moreover, both LPA and GPA bypass existing defenses, underscoring the fragility of multimodal RAG and establishing MM-PoisonRAG as a foundation for future research on securing RAG frameworks against multimodal knowledge poisoning.

cs.LG↗