arXiv subjects

Hanyu Li

Publications and source records attributed to Hanyu Li.

At least 19 recordsLinked to original sources

arXiv2026

ComplexityWorld: Benchmarking Vision-Language Models on Verifiable Visual Decision Making

Vision-language models (VLMs) have made rapid progress in visual perception and increasingly support real-world tasks that depend on images. Many such tasks, however, require more than rec- ognizing what an image contains: a model must use visual evidence to make a complete decision whose parts jointly satisfy global constraints. We introduce COMPLEXITYWORLD, a benchmark of 390 tasks across 39 domain-inspired visual worlds and 29 decision categories. Each task is generated from a hidden structured specification, rendered as a visual scene, and scored by an exe- cutable verifier that accepts any feasible solution. Under direct inference, all evaluated models ex- cept GPT-5.6-Sol remain below 40% verifier ac- ceptance rate (VAR), while GPT-5.6-Sol reaches 75.6%. Performance improves substantially when the same decision information is made explicit in structured form, yet varies sharply across equiva- lent visual presentations. Agent scaffolds provide smaller, model-dependent gains. Together, these results reveal a persistent visual-to-decision bot- tleneck that additional inference alone does not remove.

cs.CV↗

arXiv2026

Statistical equivalence of reduced gravity and enhanced friction in granular packings

Using X-ray tomography, we compare granular packings prepared under buoyancy-reduced effective gravity with normal gravity packings of particles with systematically varied friction. We show that reducing gravity lowers the random loose packing limit in a manner analogous to increasing friction. Granular packings under reduced gravity and with enhanced friction exhibit identical volume distributions, compactivity, and entropy, indicating that both routes sample statistically equivalent Edwards volume ensembles of mechanically stable states. This equivalence originates from a common relaxation of the mechanical stability constraint: under both conditions, fewer particles are required to participate in the underlying load-bearing bridge structures, leading to a lower contact-number requirement and a higher density of mechanically stable states. Nevertheless, reduced gravity retains a distinct contact-scale signature through more isotropic contact orientations. These findings identify gravity as a physical control governing the statistical accessibility of mechanically stable states within the Edwards framework and provide a unified statistical description of granular packings formed through different physical routes.

cond-mat.soft↗

arXiv2026

UniPinRec: Unifying Generative Retrieval and Ranking at Pinterest Scale

Modern recommendation systems predominantly train retrieval and ranking as separate models despite both increasingly relying on large transformers encoding the same user behavior data, duplicating parameters, compute, and serving cost. Prior work unifies the model architecture but not the full pipeline: input formats, training procedures, and serving stacks remain fragmented across stages. We present UniPinRec, which achieves full-stack unification of retrieval and ranking at Pinterest: one input format, one model, one training stage, deployed within existing serving infrastructure. A shared transformer encodes the user action sequence into candidate-independent representations that branch into retrieval (ANN dot-product) and ranking (cross-attention) via task-specific heads. Three ideas make this work: (1) Masked Action Modeling (MAM) eliminates interleaving, enabling weight sharing without doubling context length; (2) Blended training examples pair action sequences with feedview impression slates to satisfy both objectives jointly; (3) Cross-stage KV cache sharing reuses user-history computation from retrieval for ranking, reducing total FLOPs versus serving two independent models. Deployed in the Pinterest core surfaces, UniPinRec delivers approximately +1% online engagement lift while cutting end-to-end serving latency by 11.1% and lifting QPS by 63.6%. To our knowledge, this is the first full-stack unification of retrieval and ranking, covering inputs, model, training and serving, deployed in a production recommendation system.

cs.IR↗

arXiv2026

RepoMirage: Probing Repository Context Reasoning in Code Agents with Perturbations

Code agents are currently having skillful performance on repository-level software engineering benchmarks, but it remains unclear whether success on end-to-end tasks such as issue resolution truly reflects repository context reasoning, the ability to identify the task-relevant information across multiple files and reason over the relations among them. To investigate this question, we introduce RepoMirage, a two-stage evaluation suite built on SWE-Bench Verified that adopts perturbation as a diagnostic tool to increase the demand for context reasoning by transforming how the repository is exposed. First, RepoMirage-Perturb applies three types of semantics-preserving repository-level perturbations, revealing a clear performance drop when correct solving requires broader context access. RepoMirage-Extend further turns perturbation-targeted structural bottlenecks into explicit tasks beyond issue resolution, where the average performance declines from 66.8% in the original setting to 25.3%, indicating a significant deficiency in repository context reasoning. Further trajectory analysis reveals an exploration drift, where agents access broader repository context but fail to turn it into effective structure information. Motivated by this observation, we propose RepoAnchor, a structure-first prototype workflow that separates repository exploration from downstream problem solving, and show that explicit structural scaffolding yields notable gains. These results uncover an previously overlooked gap in repository context reasoning for code agents and suggest that stronger structure-aware methods are potential to improve them.

cs.SE↗

arXiv2026

An Information-Theoretic Criterion for Efficient Data Synthesis

Synthetic data becomes crucial for large language model training, but its effectiveness is highly inconsistent. We provide an information-theoretic account of this inconsistency: synthetic data improves a model only when the generation-training loop is information-open, i.e., shaped by external signals (verifiers, environments, or rubrics) that inject task-relevant information beyond the model's current distribution. When the loop is information-closed (relying on the model's own outputs without such signals), the data processing inequality ensures that task-relevant information can only decrease, making collapse a predicted outcome. Among information-open pipelines, both efficiency and generalization hinge on the meta-level of supervision: a coarser signal such as binary correctness treats all acceptable outputs as equivalent, so the behavior it teaches is not tied to any particular domain or surface form and generalizes naturally across tasks and domains. These observations lead to a guiding thesis: learning preferentially converges to the most information-efficient signal component available, which accelerates learning when that component is the intended one, but causes reward hacking when a spurious pattern happens to be simpler.

cs.LG↗

arXiv2026

A Unified Glassy Rheology for Granular Matter

Granular flows are ubiquitous in nature and industrial applications, yet a complete continuum theory remains a long-standing challenge. The leading empirical approach, {\mu}(I) rheology, lacks microscopic foundations and becomes multivalued in dense, slowly sheared flows where nonlocal corrections are required. Exploiting state-of-the-art high-speed X-ray tomography to investigate microscopic dynamics of dense granular flows in a Couette geometry, we establish a new, universal constitutive law spanning quasi-static to inertial regimes based on structural relaxation, resolving the fundamental difficulty in the original {\mu}(I) framework. By further establishing a non-equilibrium statistical framework for granular flows, we demonstrate an intrinsic analogy between driven granular matter and hard-sphere liquids owing to their identical Carnahan-Starling equation of state, naturally explaining our rheological approach and the emergence of glassy behaviors. Our framework unifies granular rheology with the broader physics of disordered systems and provides a complete, microscopically-based theoretical framework for dense granular flow.

cond-mat.soft↗

arXiv2026

SaFeR-Steer: Evolving Multi-Turn MLLMs via Synthetic Bootstrapping and Feedback Dynamics

MLLMs are increasingly deployed in multi-turn settings, where attackers can escalate unsafe intent through the evolving visual-text history and exploit long-context safety decay. Yet safety alignment is still dominated by single-turn data and fixed-template dialogues, leaving a mismatch between training and deployment. To bridge this gap, we propose SaFeR-Steer, a progressive multi-turn alignment framework that combines staged synthetic bootstrapping with tutor-in-the-loop GRPO to train a single student under adaptive, on-policy attacks. We also introduce Trajectory-Consistent Summative Reward (TCSR), which aggregates the historical minimum and average of turn rewards so that any low-quality turn affects the trajectory-level return. I. Dataset. We release STEER, a multi-turn multimodal safety dataset with STEER-SFT (12,934), STEER-RL (2,000), and STEER-Bench (3,227) dialogues spanning 2-10 turns. II. Experiment. Starting from Qwen2.5-VL-3B/7B, SaFeR-Steer substantially improves Safety/Helpfulness on both single-turn (48.30/45.86 $\rightarrow$ 81.84/70.77 for 3B; 56.21/60.32 $\rightarrow$ 87.89/77.40 for 7B) and multi-turn benchmarks (12.55/27.13 $\rightarrow$ 55.58/70.27 for 3B; 24.66/46.48 $\rightarrow$ 64.89/72.35 for 7B), shifting failures to later turns and yielding robustness beyond scaling alone. Code is available at https://anonymous.4open.science/r/SaFeR-Steer

cs.LG↗

arXiv2026

Scalable Preconditioners for the Pseudo-4D DFN Lithium-ion Battery Model

The pseudo-4D Doyle-Fuller-Newman (DFN) model enables predictive simulation of lithium-ion batteries with three-dimensional electrode architectures and particle-scale diffusion, extending the standard pseudo-2D (P2D) formulation to fully resolve cell geometry. This leads to large, nonlinear systems with strong coupling across multiple physical scales, posing significant challenges for scalable numerical solution. We introduce block-structured preconditioning strategies that exploit the mathematical properties of the coupled system, employing multigrid techniques for electrode-level operators and localized solvers for particle-scale diffusion. Comprehensive scalability studies are performed across a range of geometries, including homogeneous and heterogeneous cubic cells, flattened jelly-roll configurations, and triply periodic minimal surface electrodes, to assess solver robustness and parallel scalability. The proposed methods consistently deliver efficient convergence and enable the solution of battery models with hundreds of millions of degrees of freedom on large-scale parallel hardware.

math.NA↗

arXiv2026

JudgeRLVR: Judge First, Generate Second for Efficient Reasoning

Reinforcement Learning with Verifiable Rewards (RLVR) has become a standard paradigm for reasoning in Large Language Models. However, optimizing solely for final-answer correctness often drives models into aimless, verbose exploration, where they rely on exhaustive trial-and-error tactics rather than structured planning to reach solutions. While heuristic constraints like length penalties can reduce verbosity, they often truncate essential reasoning steps, creating a difficult trade-off between efficiency and verification. In this paper, we argue that discriminative capability is a prerequisite for efficient generation: by learning to distinguish valid solutions, a model can internalize a guidance signal that prunes the search space. We propose JudgeRLVR, a two-stage judge-then-generate paradigm. In the first stage, we train the model to judge solution responses with verifiable answers. In the second stage, we fine-tune the same model with vanilla generating RLVR initialized from the judge. Compared to Vanilla RLVR using the same math-domain training data, JudgeRLVR achieves a better quality--efficiency trade-off for Qwen3-30B-A3B: on in-domain math, it delivers about +3.7 points average accuracy gain with -42\% average generation length; on out-of-domain benchmarks, it delivers about +4.5 points average accuracy improvement, demonstrating enhanced generalization.

cs.CL↗

arXiv2026

MiMo-V2-Flash Technical Report

We present MiMo-V2-Flash, a Mixture-of-Experts (MoE) model with 309B total parameters and 15B active parameters, designed for fast, strong reasoning and agentic capabilities. MiMo-V2-Flash adopts a hybrid attention architecture that interleaves Sliding Window Attention (SWA) with global attention, with a 128-token sliding window under a 5:1 hybrid ratio. The model is pre-trained on 27 trillion tokens with Multi-Token Prediction (MTP), employing a native 32k context length and subsequently extended to 256k. To efficiently scale post-training compute, MiMo-V2-Flash introduces a novel Multi-Teacher On-Policy Distillation (MOPD) paradigm. In this framework, domain-specialized teachers (e.g., trained via large-scale reinforcement learning) provide dense and token-level reward, enabling the student model to perfectly master teacher expertise. MiMo-V2-Flash rivals top-tier open-weight models such as DeepSeek-V3.2 and Kimi-K2, despite using only 1/2 and 1/3 of their total parameters, respectively. During inference, by repurposing MTP as a draft model for speculative decoding, MiMo-V2-Flash achieves up to 3.6 acceptance length and 2.6x decoding speedup with three MTP layers. We open-source both the model weights and the three-layer MTP weights to foster open research and community collaboration.

cs.CL↗

arXiv2025

MiMo-Audio: Audio Language Models are Few-Shot Learners

Existing audio language models typically rely on task-specific fine-tuning to accomplish particular audio tasks. In contrast, humans are able to generalize to new audio tasks with only a few examples or simple instructions. GPT-3 has shown that scaling next-token prediction pretraining enables strong generalization capabilities in text, and we believe this paradigm is equally applicable to the audio domain. By scaling MiMo-Audio's pretraining data to over one hundred million of hours, we observe the emergence of few-shot learning capabilities across a diverse set of audio tasks. We develop a systematic evaluation of these capabilities and find that MiMo-Audio-7B-Base achieves SOTA performance on both speech intelligence and audio understanding benchmarks among open-source models. Beyond standard metrics, MiMo-Audio-7B-Base generalizes to tasks absent from its training data, such as voice conversion, style transfer, and speech editing. MiMo-Audio-7B-Base also demonstrates powerful speech continuation capabilities, capable of generating highly realistic talk shows, recitations, livestreaming and debates. At the post-training stage, we curate a diverse instruction-tuning corpus and introduce thinking mechanisms into both audio understanding and generation. MiMo-Audio-7B-Instruct achieves open-source SOTA on audio understanding benchmarks (MMSU, MMAU, MMAR, MMAU-Pro), spoken dialogue benchmarks (Big Bench Audio, MultiChallenge Audio) and instruct-TTS evaluations, approaching or surpassing closed-source models. Model checkpoints and full evaluation suite are available at https://github.com/XiaomiMiMo/MiMo-Audio.

cs.CL↗

arXiv2025

Reinforcement Learning for Chain of Thought Compression with One-Domain-to-All Generalization

Chain-of-thought reasoning in large language models can trigger an "overthinking trap": longer rollouts raise cost and latency yet often yield unreliable accuracy gains. Existing methods use global, static controls that may suppress needed reasoning. We propose mastery-gated, sample-level, soft reinforcement learning compression that penalizes long rollouts only when the model already solves the problem and has produced a shorter rollout. Across benchmarks, it cuts response length by 20-40% with comparable or higher accuracy and generalizes across domains: a model trained on math spontaneously shortens unseen tasks (code, instruction following, general-knowledge QA) without hurting accuracy. We further show two-way transfer between non-agent CoT and tool-use agents: non-agent training reduces SWE-Bench Verified rounds by 13%, while compressing a thinking agent cuts SWE trajectories by 67% tokens and 52% rounds and shortens non-agent outputs by up to 44%. Compression is thus not cosmetic brevity, but an inherent computation policy -- what to keep, and what to forget.

cs.CL↗

arXiv2025

Will AI Trade? A Computational Inversion of the No-Trade Theorem

Classic no-trade theorems attribute trade to heterogeneous beliefs. We re-examine this conclusion for AI agents, asking if trade can arise from computational limitations, under common beliefs. We model agents' bounded computational rationality within an unfolding game framework, where computational power determines the complexity of its strategy. Our central finding inverts the classic paradigm: a stable no-trade outcome (Nash equilibrium) is reached only when "almost rational" agents have slightly different computational power. Paradoxically, when agents possess identical power, they may fail to converge to equilibrium, resulting in persistent strategic adjustments that constitute a form of trade. This instability is exacerbated if agents can strategically under-utilize their computational resources, which eliminates any chance of equilibrium in Matching Pennies scenarios. Our results suggest that the inherent computational limitations of AI agents can lead to situations where equilibrium is not reached, creating a more lively and unpredictable trade environment than traditional models would predict.

cs.GT↗

arXiv2025

SAIL-Embedding Technical Report: Omni-modal Embedding Foundation Model

Multimodal embedding models aim to yield informative unified representations that empower diverse cross-modal tasks. Despite promising developments in the evolution from CLIP-based dual-tower architectures to large vision-language models, prior works still face unavoidable challenges in real-world applications and business scenarios, such as the limited modality support, unstable training mechanisms, and industrial domain gaps. In this work, we introduce SAIL-Embedding, an omni-modal embedding foundation model that addresses these issues through tailored training strategies and architectural design. In the optimization procedure, we propose a multi-stage training scheme to boost the multifaceted effectiveness of representation learning. Specifically, the content-aware progressive training aims to enhance the model's adaptability to diverse downstream tasks and master enriched cross-modal proficiency. The collaboration-aware recommendation enhancement training further adapts multimodal representations for recommendation scenarios by distilling knowledge from sequence-to-item and ID-to-item embeddings while mining user historical interests. Concurrently, we develop the stochastic specialization and dataset-driven pattern matching to strengthen model training flexibility and generalizability. Experimental results show that SAIL-Embedding achieves SOTA performance compared to other methods in different retrieval tasks. In online experiments across various real-world scenarios integrated with our model, we observe a significant increase in Lifetime (LT), which is a crucial indicator for the recommendation experience. For instance, the model delivers the 7-day LT gain of +0.5% in the Douyin-Selected scenario. For the Douyin feed rank model, the match features produced by SAIL-Embedding yield a +0.1% AUC gain.

cs.IR↗

arXiv2025

How Large Language Models Need Symbolism

We argue that AI's future requires more than scaling. To unlock genuine discovery, large language models need a compass: human-crafted symbols to guide their powerful but blind intuition.

cs.CL↗

arXiv2025

Discovering Expert-Level Nash Equilibrium Algorithms with Large Language Models

Designing polynomial-time algorithms for approximate Nash equilibria (ANE) with provable worst-case guarantees is a fundamental open problem in algorithmic game theory. While large language models (LLMs) can generate candidate algorithms at scale, certifying worst-case guarantees requires formal analysis over all game instances -- a task for which no automated system previously existed. Here, we present LegoNE, a framework encoding expert proof strategies into a symbolic language that automatically compiles any candidate algorithm into a finite optimization problem certifying its worst-case guarantee. Integrating LegoNE with a reasoning LLM, we rediscovered an algorithm matching the best polynomial-time guarantee for two-player games, and discovered a three-player algorithm improving the best guarantee from $0.6+\delta$ to $0.5+\delta$ -- provably beyond the reach of the extension technique, the only previously known multi-player ANE design paradigm. These results show that encoding domain-specific proof strategies into a machine-tractable language can support LLM-driven discovery of algorithms outside known human design paradigms.

cs.GT↗

arXiv2025

PinFM: Foundation Model for User Activity Sequences at a Billion-scale Visual Discovery Platform

User activity sequences have emerged as one of the most important signals in recommender systems. We present a foundational model, PinFM, for understanding user activity sequences across multiple applications at a billion-scale visual discovery platform. We pretrain a transformer model with 20B+ parameters using extensive user activity data, then fine-tune it for specific applications, efficiently coupling it with existing models. While this pretraining-and-fine-tuning approach has been popular in other domains, such as Vision and NLP, its application in industrial recommender systems presents numerous challenges. The foundational model must be scalable enough to score millions of items every second while meeting tight cost and latency constraints imposed by these systems. Additionally, it should capture the interactions between user activities and other features and handle new items that were not present during the pretraining stage. We developed innovative techniques to address these challenges. Our infrastructure and algorithmic optimizations, such as the Deduplicated Cross-Attention Transformer (DCAT), improved our throughput by 600% on Pinterest internal data. We demonstrate that PinFM can learn interactions between user sequences and candidate items by altering input sequences, leading to a 20% increase in engagement with new items. PinFM is now deployed to help improve the experience of more than half a billion users across various applications.

cs.LG↗

arXiv2025

IDA-Bench: Evaluating LLMs on Interactive Guided Data Analysis

Large Language Models (LLMs) show promise as data analysis agents, but existing benchmarks overlook the iterative nature of the field, where experts' decisions evolve with deeper insights of the dataset. To address this, we introduce IDA-Bench, a novel benchmark evaluating LLM agents in multi-round interactive scenarios. Derived from complex Kaggle notebooks, tasks are presented as sequential natural language instructions by an LLM-simulated user. Agent performance is judged by comparing its final numerical output to the human-derived baseline. Initial results show that even state-of-the-art coding agents (like Claude-3.7-thinking) succeed on < 50% of the tasks, highlighting limitations not evident in single-turn tests. This work underscores the need to improve LLMs' multi-round capabilities for building more reliable data analysis agents, highlighting the necessity of achieving a balance between instruction following and reasoning.

cs.CL↗