Zhiyuan Hu

Lin Xu, Zhiyuan Hu, Daquan Zhou et al. · berkeley

Large Language Models (LLMs) have significantly advanced natural language processing, demonstrating exceptional reasoning, tool usage, and memory capabilities. As their applications expand into multi-agent environments, there arises a need for a comprehensive evaluation framework that captures LLMs' reasoning, planning, collaboration, and other social abilities. This work introduces a novel competition-based benchmark framework specifically designed to assess LLMs within multi-agent settings, providing quantitative metrics to evaluate their judgment, reasoning, deception, self-awareness, cooperation, coordination, and rationality. We utilize two social deduction games alongside three game-theory scenarios to create diverse environments. Our frame is fortified with the probabilistic graphic modeling (PGM) method, enhancing the LLMs' capabilities in navigating complex social and cognitive dimensions. We evaluate seven LLMs, quantitatively highlighting a significant capability gap of over threefold between the strongest, GPT o1, and the weakest, Llama-2-70B. It also confirms that our PGM enhancement boosts the abilities of all selected models by an average of 37%. Our data and code can be found here https://github.com/cathyxl/MAgIC.

Zhiyuan Hu, Yuliang Liu, Jinman Zhao et al.

Large language models (LLMs) face significant challenges in handling long-context tasks because of their limited effective context window size during pretraining, which restricts their ability to generalize over extended sequences. Meanwhile, extending the context window in LLMs through post-pretraining is highly resource-intensive. To address this, we introduce LongRecipe, an efficient training strategy for extending the context window of LLMs, including impactful token analysis, position index transformation, and training optimization strategies. It simulates long-sequence inputs while maintaining training efficiency and significantly improves the model's understanding of long-range dependencies. Experiments on three types of LLMs show that LongRecipe can utilize long sequences while requiring only 30% of the target context window size, and reduces computational training resource over 85% compared to full sequence training. Furthermore, LongRecipe also preserves the original LLM's capabilities in general tasks. Ultimately, we can extend the effective context window of open-source LLMs from 8k to 128k, achieving performance close to GPT-4 with just one day of dedicated training using a single GPU with 80G memory. Our code is released at https://github.com/zhiyuanhubj/LongRecipe.

Thong Nguyen, Yi Bin, Xiaobao Wu et al. · mit

Data quality stands at the forefront of deciding the effectiveness of video-language representation learning. However, video-text pairs in previous data typically do not align perfectly with each other, which might lead to video-language representations that do not accurately reflect cross-modal semantics. Moreover, previous data also possess an uneven distribution of concepts, thereby hampering the downstream performance across unpopular subjects. To address these problems, we propose MAMA, a new approach to learning video-language representations by utilizing a contrastive objective with a subtractive angular margin to regularize cross-modal representations in their effort to reach perfect similarity. Furthermore, to adapt to the non-uniform concept distribution, MAMA utilizes a multi-layer perceptron (MLP)-parameterized weighting function that maps loss values to sample weights which enable dynamic adjustment of the model's focus throughout the training. With the training guided by a small amount of unbiased meta-data and augmented by video-text data generated by large vision-language model, MAMA improves video-language representations and achieve superior performances on commonly used video question answering and text-video retrieval datasets. The code, model, and data have been made available at https://nguyentthong.github.io/MAMA.

Zhiyuan Hu, Yunsheng Li, Jiancheng Lyu et al.

The problem of class incremental learning (CIL) is considered. State-of-the-art approaches use a dynamic architecture based on network expansion (NE), in which a task expert is added per task. While effective from a computational standpoint, these methods lead to models that grow quickly with the number of tasks. A new NE method, dense network expansion (DNE), is proposed to achieve a better trade-off between accuracy and model complexity. This is accomplished by the introduction of dense connections between the intermediate layers of the task expert networks, that enable the transfer of knowledge from old to new tasks via feature sharing and reusing. This sharing is implemented with a cross-task attention mechanism, based on a new task attention block (TAB), that fuses information across tasks. Unlike traditional attention mechanisms, TAB operates at the level of the feature mixing and is decoupled with spatial attentions. This is shown more effective than a joint spatial-and-task attention for CIL. The proposed DNE approach can strictly maintain the feature space of old classes while growing the network and feature scale at a much slower rate than previous methods. In result, it outperforms the previous SOTA methods by a margin of 4\% in terms of accuracy, with similar or even smaller model scale.

Puze Liu, Kuo Zhang, Davide Tateo et al.

Safety is a crucial property of every robotic platform: any control policy should always comply with actuator limits and avoid collisions with the environment and humans. In reinforcement learning, safety is even more fundamental for exploring an environment without causing any damage. While there are many proposed solutions to the safe exploration problem, only a few of them can deal with the complexity of the real world. This paper introduces a new formulation of safe exploration for reinforcement learning of various robotic tasks. Our approach applies to a wide class of robotic platforms and enforces safety even under complex collision constraints learned from data by exploring the tangent space of the constraint manifold. Our proposed approach achieves state-of-the-art performance in simulated high-dimensional and dynamic tasks while avoiding collisions with the environment. We show safe real-world deployment of our learned controller on a TIAGo++ robot, achieving remarkable performance in manipulation and human-robot interaction tasks.

Xiangyan Liu, Bo Lan, Zhiyuan Hu et al.

Large Language Models (LLMs) excel in stand-alone code tasks like HumanEval and MBPP, but struggle with handling entire code repositories. This challenge has prompted research on enhancing LLM-codebase interaction at a repository scale. Current solutions rely on similarity-based retrieval or manual tools and APIs, each with notable drawbacks. Similarity-based retrieval often has low recall in complex tasks, while manual tools and APIs are typically task-specific and require expert knowledge, reducing their generalizability across diverse code tasks and real-world applications. To mitigate these limitations, we introduce CodexGraph, a system that integrates LLM agents with graph database interfaces extracted from code repositories. By leveraging the structural properties of graph databases and the flexibility of the graph query language, CodexGraph enables the LLM agent to construct and execute queries, allowing for precise, code structure-aware context retrieval and code navigation. We assess CodexGraph using three benchmarks: CrossCodeEval, SWE-bench, and EvoCodeBench. Additionally, we develop five real-world coding applications. With a unified graph database schema, CodexGraph demonstrates competitive performance and potential in both academic and real-world environments, showcasing its versatility and efficacy in software engineering. Our application demo: https://github.com/modelscope/modelscope-agent/tree/master/apps/codexgraph_agent.

Zhiyuan Hu, Jiancheng Lyu, Dashan Gao et al.

Continual learning (CL) has attracted increasing attention in the recent past. It aims to mimic the human ability to learn new concepts without catastrophic forgetting. While existing CL methods accomplish this to some extent, they are still prone to semantic drift of the learned feature space. Foundation models, which are endowed with a robust feature representation, learned from very large datasets, provide an interesting substrate for the solution of the CL problem. Recent work has also shown that they can be adapted to specific tasks by prompt tuning techniques that leave the generality of the representation mostly unscathed. An open question is, however, how to learn both prompts that are task specific and prompts that are global, i.e. capture cross-task information. In this work, we propose the Prompt Of Prompts (POP) model, which addresses this goal by progressively learning a group of task-specified prompts and a group of global prompts, denoted as POP, to integrate information from the former. We show that a foundation model equipped with POP learning is able to outperform classic CL methods by a significant margin. Moreover, as prompt tuning only requires a small set of training samples, POP is able to perform CL in the few-shot setting, while still outperforming competing methods trained on the entire dataset.

Miao Xiong, Zhiyuan Hu, Xinyang Lu et al.

Empowering large language models to accurately express confidence in their answers is essential for trustworthy decision-making. Previous confidence elicitation methods, which primarily rely on white-box access to internal model information or model fine-tuning, have become less suitable for LLMs, especially closed-source commercial APIs. This leads to a growing need to explore the untapped area of black-box approaches for LLM uncertainty estimation. To better break down the problem, we define a systematic framework with three components: prompting strategies for eliciting verbalized confidence, sampling methods for generating multiple responses, and aggregation techniques for computing consistency. We then benchmark these methods on two key tasks-confidence calibration and failure prediction-across five types of datasets (e.g., commonsense and arithmetic reasoning) and five widely-used LLMs including GPT-4 and LLaMA 2 Chat. Our analysis uncovers several key insights: 1) LLMs, when verbalizing their confidence, tend to be overconfident, potentially imitating human patterns of expressing confidence. 2) As model capability scales up, both calibration and failure prediction performance improve. 3) Employing our proposed strategies, such as human-inspired prompts, consistency among multiple responses, and better aggregation strategies can help mitigate this overconfidence from various perspectives. 4) Comparisons with white-box methods indicate that while white-box methods perform better, the gap is narrow, e.g., 0.522 to 0.605 in AUROC. Despite these advancements, none of these techniques consistently outperform others, and all investigated methods struggle in challenging tasks, such as those requiring professional knowledge, indicating significant scope for improvement. We believe this study can serve as a strong baseline and provide insights for eliciting confidence in black-box LLMs.

Zhiyuan Hu, Yue Feng, Anh Tuan Luu et al.

Dialogue systems and large language models (LLMs) have gained considerable attention. However, the direct utilization of LLMs as task-oriented dialogue (TOD) models has been found to underperform compared to smaller task-specific models. Nonetheless, it is crucial to acknowledge the significant potential of LLMs and explore improved approaches for leveraging their impressive abilities. Motivated by the goal of leveraging LLMs, we propose an alternative approach called User-Guided Response Optimization (UGRO) to combine it with a smaller TOD model. This approach uses LLM as annotation-free user simulator to assess dialogue responses, combining them with smaller fine-tuned end-to-end TOD models. By utilizing the satisfaction feedback generated by LLMs, UGRO further optimizes the supervised fine-tuned TOD model. Specifically, the TOD model takes the dialogue history as input and, with the assistance of the user simulator's feedback, generates high-satisfaction responses that meet the user's requirements. Through empirical experiments on two TOD benchmarks, we validate the effectiveness of our method. The results demonstrate that our approach outperforms previous state-of-the-art (SOTA) results.

Yuzhao Chen, Zonghuan Li, Zhiyuan Hu et al.

The problem of continual learning has attracted rising attention in recent years. However, few works have questioned the commonly used learning setup, based on a task curriculum of random class. This differs significantly from human continual learning, which is guided by taxonomic curricula. In this work, we propose the Taxonomic Class Incremental Learning (TCIL) problem. In TCIL, the task sequence is organized based on a taxonomic class tree. We unify existing approaches to CIL and taxonomic learning as parameter inheritance schemes and introduce a new such scheme for the TCIL learning. This enables the incremental transfer of knowledge from ancestor to descendant class of a class taxonomy through parameter inheritance. Experiments on CIFAR-100 and ImageNet-100 show the effectiveness of the proposed TCIL method, which outperforms existing SOTA methods by 2% in terms of final accuracy on CIFAR-100 and 3% on ImageNet-100.

Yue Liu, Zhiyuan Hu, Flood Sung et al.

This paper introduces KLong, an open-source LLM agent trained to solve extremely long-horizon tasks. The principle is to first cold-start the model via trajectory-splitting SFT, then scale it via progressive RL training. Specifically, we first activate basic agentic abilities of a base model with a comprehensive SFT recipe. Then, we introduce Research-Factory, an automated pipeline that generates high-quality training data by collecting research papers and constructing evaluation rubrics. Using this pipeline, we build thousands of long-horizon trajectories distilled from Claude 4.5 Sonnet (Thinking). To train with these extremely long trajectories, we propose a new trajectory-splitting SFT, which preserves early context, progressively truncates later context, and maintains overlap between sub-trajectories. In addition, to further improve long-horizon task-solving capability, we propose a novel progressive RL, which schedules training into multiple stages with progressively extended timeouts. Experiments demonstrate the superiority and generalization of KLong, as shown in Figure 1. Notably, our proposed KLong (106B) surpasses Kimi K2 Thinking (1T) by 11.28% on PaperBench, and the performance improvement generalizes to other coding benchmarks like SWE-bench Verified and MLE-bench.

Zhiyuan Hu, Yue Feng, Yang Deng et al.

Recently, the development of large language models (LLMs) has been significantly enhanced the question answering and dialogue generation, and makes them become increasingly popular in current practical scenarios. While unlike the general dialogue system which emphasizes the semantic performance, the task-oriented dialogue (ToD) systems aim to achieve the dialogue goal efficiently and successfully in multiple turns. Unfortunately, existing LLM-induced ToD systems lack the direct reward toward the final goal and do not take account of the dialogue proactivity that can strengthen the dialogue efficiency. To fill these gaps, we introduce the ProToD (Proactively Goal-Driven LLM-Induced ToD) approach, which anticipates the future dialogue actions and incorporates the goal-oriented reward signal to enhance ToD systems. Additionally, we present a novel evaluation method that assesses ToD systems based on goal-driven dialogue simulations. This method allows us to gauge user satisfaction, system efficiency and successful rate while overcoming the limitations of current Information and Success metrics. Empirical experiments conducted on the MultiWoZ 2.1 dataset demonstrate that our model can achieve superior performance using only 10% of the data compared to previous end-to-end fully supervised models. This improvement is accompanied by enhanced user satisfaction and efficiency.

Yufei He, Juncheng Liu, Zhiyuan Hu et al.

Prevailing medical AI operates on an unrealistic ''one-shot'' model, diagnosing from a complete patient file. However, real-world diagnosis is an iterative inquiry where Clinicians sequentially ask questions and order tests to strategically gather information while managing cost and time. To address this, we first propose Med-Inquire, a new benchmark designed to evaluate an agent's ability to perform multi-turn diagnosis. Built upon a dataset of real-world clinical cases, Med-Inquire simulates the diagnostic process by hiding a complete patient file behind specialized Patient and Examination agents. They force the agent to proactively ask questions and order tests to gather information piece by piece. To tackle the challenges posed by Med-Inquire, we then introduce EvoClinician, a self-evolving agent that learns efficient diagnostic strategies at test time. Its core is a ''Diagnose-Grade-Evolve'' loop: an Actor agent attempts a diagnosis; a Process Grader agent performs credit assignment by evaluating each action for both clinical yield and resource efficiency; finally, an Evolver agent uses this feedback to update the Actor's strategy by evolving its prompt and memory. Our experiments show EvoClinician outperforms continual learning baselines and other self-evolving agents like memory agents. The code is available at https://github.com/yf-he/EvoClinician

Zhiyuan Hu, Yunhai Hu, Juncheng Liu et al.

Multi-agent systems have evolved into practical LLM-driven collaborators for many applications, gaining robustness from diversity and cross-checking. However, multi-agent RL (MARL) training is resource-intensive and unstable: co-adapting teammates induce non-stationarity, and rewards are often sparse and high-variance. Therefore, we introduce \textbf{Multi-Agent Test-Time Reinforcement Learning (MATTRL)}, a framework that injects structured textual experience into multi-agent deliberation at inference time. MATTRL forms a multi-expert team of specialists for multi-turn discussions, retrieves and integrates test-time experiences, and reaches consensus for final decision-making. We also study credit assignment for constructing a turn-level experience pool, then reinjecting it into the dialogue. Across challenging benchmarks in medicine, math, and education, MATTRL improves accuracy by an average of 3.67\% over a multi-agent baseline, and by 8.67\% over comparable single-agent baselines. Ablation studies examine different credit-assignment schemes and provide a detailed comparison of how they affect training outcomes. MATTRL offers a stable, effective and efficient path to distribution-shift-robust multi-agent reasoning without tuning.

Zhiyuan Hu, Yucheng Wang, Yufei He et al.

Reinforcement learning (RL) has become a central paradigm for post-training large language models (LLMs), particularly for complex reasoning tasks, yet it often suffers from exploration collapse: policies prematurely concentrate on a small set of dominant reasoning patterns, improving pass@1 while limiting rollout-level diversity and gains in pass@k. We argue that this failure stems from regularizing local token behavior rather than diversity over sets of solutions. To address this, we propose Uniqueness-Aware Reinforcement Learning, a rollout-level objective that explicitly rewards correct solutions that exhibit rare high-level strategies. Our method uses an LLM-based judge to cluster rollouts for the same problem according to their high-level solution strategies, ignoring superficial variations, and reweights policy advantages inversely with cluster size. As a result, correct but novel strategies receive higher rewards than redundant ones. Across mathematics, physics, and medical reasoning benchmarks, our approach consistently improves pass@$k$ across large sampling budgets and increases the area under the pass@$k$ curve (AUC@$K$) without sacrificing pass@1, while sustaining exploration and uncovering more diverse solution strategies at scale.

Zhiyuan Hu, Chumin Liu, Yue Feng et al.

Controllable text generation is a challenging and meaningful field in natural language generation (NLG). Especially, poetry generation is a typical one with well-defined and strict conditions for text generation which is an ideal playground for the assessment of current methodologies. While prior works succeeded in controlling either semantic or metrical aspects of poetry generation, simultaneously addressing both remains a challenge. In this paper, we pioneer the use of the Diffusion model for generating sonnets and Chinese SongCi poetry to tackle such challenges. In terms of semantics, our PoetryDiffusion model, built upon the Diffusion model, generates entire sentences or poetry by comprehensively considering the entirety of sentence information. This approach enhances semantic expression, distinguishing it from autoregressive and large language models (LLMs). For metrical control, the separation feature of diffusion generation and its constraint control module enable us to flexibly incorporate a novel metrical controller to manipulate and evaluate metrics (format and rhythm). The denoising process in PoetryDiffusion allows for gradual enhancement of semantics and flexible integration of the metrical controller which can calculate and impose penalties on states that stray significantly from the target control distribution. Experimental results on two datasets demonstrate that our model outperforms existing models in automatic evaluation of semantic, metrical, and overall performance as well as human evaluation.

Zhiyuan Hu, Chumin Liu, Xidong Feng et al.

In the face of uncertainty, the ability to *seek information* is of fundamental importance. In many practical applications, such as medical diagnosis and troubleshooting, the information needed to solve the task is not initially given and has to be actively sought by asking follow-up questions (for example, a doctor asking a patient for more details about their symptoms). In this work, we introduce Uncertainty of Thoughts (UoT), an algorithm to augment large language models with the ability to actively seek information by asking effective questions. UoT combines 1) an *uncertainty-aware simulation approach* which enables the model to simulate possible future scenarios and how likely they are to occur, 2) *uncertainty-based rewards* motivated by information gain which incentivizes the model to seek information, and 3) a *reward propagation scheme* to select the optimal question to ask in a way that maximizes the expected reward. In experiments on medical diagnosis, troubleshooting, and the `20 Questions` game, UoT achieves an average performance improvement of 38.1% in the rate of successful task completion across multiple LLMs compared with direct prompting and also improves efficiency (i.e., the number of questions needed to complete the task). Our code has been released [here](https://github.com/zhiyuanhubj/UoT)

Yibo Li, Jiashuo Yang, Zhi Zheng et al.

LLM agents have shown strong performance across a wide range of complex tasks, including interactive environments that require long-horizon decision making. But these agents cannot learn on the fly at test time. Self-evolving agents address this by accumulating memory and reflection across episodes rather than requiring model-weight updates. However, these agents often suffer from exploration collapse: as memory grows, behavior concentrates around familiar high-reward routines, reducing the chance of discovering better alternatives. To address this problem, we propose Autonomous Policy EXploration (APEX), which builds and maintains an explicit strategy space through a strategy map-a directed acyclic graph of milestones with prerequisite dependency edges. In APEX, Fork Discovery expands the map with evidence-grounded unexplored directions, while Policy Selection balances exploration and exploitation during planning. Evaluated on nine Jericho text-adventure games and WebArena, a realistic web interaction benchmark, APEX outperforms all baselines. Extensive ablations validate each component's contribution and demonstrate robustness across diverse settings, demonstrating APEX's effectiveness for sustained exploration in self-evolving agents.

Zhiyuan Hu, Yibo Wang, Hanze Dong et al. · salesforce

Large reasoning models (LRMs) already possess a latent capacity for long chain-of-thought reasoning. Prior work has shown that outcome-based reinforcement learning (RL) can incidentally elicit advanced reasoning behaviors such as self-correction, backtracking, and verification phenomena often referred to as the model's "aha moment". However, the timing and consistency of these emergent behaviors remain unpredictable and uncontrollable, limiting the scalability and reliability of LRMs' reasoning capabilities. To address these limitations, we move beyond reliance on prompts and coincidental "aha moments". Instead, we explicitly align models with three meta-abilities: deduction, induction, and abduction, using automatically generated, self-verifiable tasks. Our three stage-pipeline individual alignment, parameter-space merging, and domain-specific reinforcement learning, boosting performance by over 10\% relative to instruction-tuned baselines. Furthermore, domain-specific RL from the aligned checkpoint yields an additional gain in performance ceiling for both 7B and 32B models across math, coding, and science benchmarks, demonstrating that explicit meta-ability alignment offers a scalable and dependable foundation for reasoning. Code is available at: https://github.com/zhiyuanhubj/Meta-Ability-Alignment

Zhiyuan Hu, Zheng Sun, Yi Wei et al.

The performance of image generation has been significantly improved in recent years. However, the study of image screening is rare and its performance with Multimodal Large Language Models (MLLMs) is unsatisfactory due to the lack of data and the weak image aesthetic reasoning ability in MLLMs. In this work, we propose a complete solution to address these problems in terms of data and methodology. For data, we collect a comprehensive image screening dataset with over 128k samples, about 640k images. Each sample consists of an original image, four generated images. The dataset evaluates the image aesthetic reasoning ability under four aspects: appearance deformation, physical shadow, placement layout, and extension rationality. Regarding data annotation, we investigate multiple approaches, including purely manual, fully automated, and answer-driven annotations, to acquire high-quality chains of thought (CoT) data in the most cost-effective manner. Methodologically, we introduce a Hard Cases Mining (HCM) strategy with a Dynamic Proportional Accuracy (DPA) reward into the Group Relative Policy Optimization (GRPO) framework, called HCM-GRPO. This enhanced method demonstrates superior image aesthetic reasoning capabilities compared to the original GRPO. Our experimental results reveal that even state-of-the-art closed-source MLLMs, such as GPT4o and Qwen-VL-Max, exhibit performance akin to random guessing in image aesthetic reasoning. In contrast, by leveraging the HCM-GRPO, we are able to surpass the scores of both large-scale open-source and leading closed-source models with a much smaller model.

Zhiyuan Hu, Shiyun Xiong, Yifan Zhang et al.

Recent advancements in visual language models (VLMs) have notably enhanced their capabilities in handling complex Graphical User Interface (GUI) interaction tasks. Despite these improvements, current frameworks often struggle to generate correct actions in challenging GUI environments. State-of-the-art commercial VLMs are black-boxes, and fine-tuning open-source VLMs for GUI tasks requires significant resources. Additionally, existing trajectory-level evaluation and refinement techniques frequently fall short due to delayed feedback and local optimization issues. To address these challenges, we propose an approach that guides VLM agents with process supervision by a reward model during GUI navigation and control at inference time. This guidance allows the VLM agent to optimize actions at each inference step, thereby improving performance in both static and dynamic environments. In particular, our method demonstrates significant performance gains in three GUI navigation tasks, achieving a 3.4% improvement in single step action accuracy for static environments, along with a around 33% increase in task success rate in one dynamic environment. With further integration of trajectory reflection and retry mechanisms, we also demonstrate even greater enhancement in task success.

Peiming Li, Zhiyuan Hu, Yang Tang et al.

Personalized alignment is crucial for enabling Large Language Models (LLMs) to engage effectively in user-centric interactions. However, current methods face a dual challenge: they fail to infer users' deep implicit preferences (including unstated goals, semantic context and risk tolerances), and they lack the defensive reasoning required to navigate real-world ambiguity. This cognitive gap leads to responses that are superficial, brittle and short-sighted. To address this, we propose Critique-Driven Reasoning Alignment (CDRA), which reframes alignment from a scalar reward-matching task into a structured reasoning process. First, to bridge the preference inference gap, we introduce the DeepPref benchmark. This dataset, comprising 3000 preference-query pairs across 20 topics, is curated by simulating a multi-faceted cognitive council that produces critique-annotated reasoning chains to deconstruct query semantics and reveal latent risks. Second, to instill defensive reasoning, we introduce the Personalized Generative Process Reward Model (Pers-GenPRM), which frames reward modeling as a personalized reasoning task. It generates a critique chain to evaluate a response's alignment with user preferences before outputting a final score based on this rationale. Ultimately, this interpretable, structured reward signal guides policy model through Critique-Driven Policy Alignment, a process-level online reinforcement learning algorithm integrating both numerical and natural language feedback. Experiments demonstrate that CDRA excels at discovering and aligning with users' true preferences while executing robust reasoning. Our code and dataset are available at https://github.com/Zephyrian-Hugh/Deep-pref.

Yangchen Zeng, Hao Peng, Rongfeng Guo et al.

We introduce TriAlignGR, a unified multitask-multimodal framework for generative recommendation that establishes two-stage multimodal semantic propagation: (i) encoding visual semantics directly into SIDs via multimodal embeddings, and (ii) enabling the model to decode these semantics through visual description tasks. Existing Semantic ID (SID) pipelines suffer from two fundamental but underexplored problems: \textbf{SID Content Degradation (SCD)}, where cascaded encoding and residual quantization discard critical multimodal and interest-level semantics; and \textbf{SID Semantic Opacity (SSO)}, where models autoregressively generate SID sequences without truly comprehending their underlying meaning, leading to hallucination and poor generalization. Prior work addresses at most text-SID alignment, leaving visual semantics and latent user interests entirely unexploited. TriAlignGR resolves both problems through three tightly integrated components: (1)~\textbf{Cross-Modal Semantic Alignment (CMSA)} integrates visual content into SID construction through both VLM-generated textual descriptions and a multimodal embedding model that directly encodes image features alongside text, ensuring that SIDs inherently carry multimodal semantics; (2)~\textbf{Multimodal Deep Interest Mining (MDIM)} leverages LLM Chain-of-Thought reasoning to extract latent user intents (\eg ``productivity-focused lifestyle'' from noise-canceling headphones) beyond surface attributes, enriching SID semantics before discretization; and (3)~\textbf{Triangular Multitask (TMT)} jointly trains on eight complementary generation tasks under a single autoregressive loss -- including two novel visual-semantic tasks (VisDesc$\to$SID, VisDesc$\to$Title) that map VLM-generated image descriptions to SIDs and titles, completing the SID-Text-Image triangle -- without requiring task-specific towers or complex loss weighting.

Yilun Zhao, Lujing Xie, Haowei Zhang et al.

We introduce MMVU, a comprehensive expert-level, multi-discipline benchmark for evaluating foundation models in video understanding. MMVU includes 3,000 expert-annotated questions spanning 27 subjects across four core disciplines: Science, Healthcare, Humanities & Social Sciences, and Engineering. Compared to prior benchmarks, MMVU features three key advancements. First, it challenges models to apply domain-specific knowledge and perform expert-level reasoning to analyze specialized-domain videos, moving beyond the basic visual perception typically assessed in current video benchmarks. Second, each example is annotated by human experts from scratch. We implement strict data quality controls to ensure the high quality of the dataset. Finally, each example is enriched with expert-annotated reasoning rationals and relevant domain knowledge, facilitating in-depth analysis. We conduct an extensive evaluation of 32 frontier multimodal foundation models on MMVU. The latest System-2-capable models, o1 and Gemini 2.0 Flash Thinking, achieve the highest performance among the tested models. However, they still fall short of matching human expertise. Through in-depth error analyses and case studies, we offer actionable insights for future advancements in expert-level, knowledge-intensive video understanding for specialized domains.

Nuo Chen, Zhiyuan Hu, Qingyun Zou et al.

Large Language Models (LLMs) are increasingly adopted as evaluators, offering a scalable alternative to human annotation. However, existing supervised fine-tuning (SFT) approaches often fall short in domains that demand complex reasoning. Judgment is inherently reasoning-intensive: beyond surface-level scoring, it requires verifying evidence, identifying errors, and justifying decisions. Through the analysis of evaluation tasks, we find a negative correlation between SFT performance gains and the proportion of reasoning-demanding samples, revealing the limits of SFT in such scenarios. To address this, we introduce JudgeLRM, a family of judgment-oriented LLMs, trained using reinforcement learning (RL) with judge-wise, outcome-driven rewards to activate reasoning capabilities. JudgeLRM consistently outperform SFT-tuned baselines in the same size, as well as other RL and SFT variants, and even surpass state-of-the-art reasoning models: notably, JudgeLRM-3B/4B exceeds GPT-4, while JudgeLRM-7B/8B/14B outperforms DeepSeek-R1 by over 2% in F1 score, with particularly strong gains on reasoning-heavy tasks. Our findings underscore the value of RL in unlocking reasoning-aligned LLM judges.

Yan Yang, Dongxu Li, Yutong Dai et al.

Graphical user interface (GUI) agents autonomously complete tasks across platforms (\eg, Linux) by sequentially decomposing user instructions into action proposals that iteratively interact with visual elements in the evolving environment. However, two main challenges arise: i) planning (\ie, the action proposal sequence) under expansive action space, where selecting an appropriate plan is non-trivial, as many valid ones may exist; ii) accurately grounding actions in complex and high-resolution interfaces, \ie, precisely interacting with visual targets. This paper investigates the aforementioned challenges with our \textbf{G}UI \textbf{T}est-time Scaling \textbf{A}gent, namely GTA1. First, we conduct test-time scaling to select the most appropriate action proposal: at each step, multiple candidate proposals are sampled and evaluated and selected by a judge model. It trades off computation for better decision quality by concurrent sampling. Second, we propose a model that improves grounding of the selected action proposals to its corresponding visual elements. Our key insight is that reinforcement learning (RL) facilitates grounding through inherent objective alignments, rewarding successful clicks on interface elements. Experimentally, GTA1 achieves state-of-the-art performance on both grounding and agent task execution benchmarks. The code and models are released here.

Xidong Feng, Bo Liu, Yan Song et al. · cmu

Artificial intelligence progresses towards the "Era of Experience," where agents are expected to learn from continuous, grounded interaction. We argue that traditional Reinforcement Learning (RL), which typically represents value as a scalar, can restrict agent's deep understanding of environments and hinders the active, deliberative learning crucial for navigating this new paradigm. To address the issue, we introduce Natural Language Reinforcement Learning (NLRL), a framework that extends RL principles into natural language counterparts. Central to NLRL is the Language Value Function (LVF), which redefines value as an interpretable linguistic narrative articulating the rationale behind an evaluation. NLRL further extends this concept to core RL components, including policy, the Bellman equation, and policy iteration. Leveraging recent advancements in Large Language Models (LLMs), NLRL can be practically implemented to achieve RL-like policy and value training through unsupervised environment interactions. Experiments over 4 multi-step agentic tasks demonstrate NLRL's effectiveness, efficiency, and its potential to foster deeper understanding and more active learning strategies.

Thong Thanh Nguyen, Yi Bin, Xiaobao Wu et al.

Temporal grounding, which localizes video moments related to a natural language query, is a core problem of vision-language learning and video understanding. To encode video moments of varying lengths, recent methods employ a multi-level structure known as a feature pyramid. In this structure, lower levels concentrate on short-range video moments, while higher levels address long-range moments. Because higher levels experience downsampling to accommodate increasing moment length, their capacity to capture information is reduced and consequently leads to degraded information in moment representations. To resolve this problem, we propose a contrastive learning framework to capture salient semantics among video moments. Our key methodology is to leverage samples from the feature space emanating from multiple stages of the video encoder itself requiring neither data augmentation nor online memory banks to obtain positive and negative samples. To enable such an extension, we introduce a sampling process to draw multiple video moments corresponding to a common query. Subsequently, by utilizing these moments' representations across video encoder layers, we instantiate a novel form of multi-scale and cross-scale contrastive learning that links local short-range video moments with global long-range video moments. Extensive experiments demonstrate the effectiveness of our framework for not only long-form but also short-form video grounding.

Thong Nguyen, Xiaobao Wu, Xinshuai Dong et al. · mit

Fully fine-tuning pretrained large-scale transformer models has become a popular paradigm for video-language modeling tasks, such as temporal language grounding and video-language summarization. With a growing number of tasks and limited training data, such full fine-tuning approach leads to costly model storage and unstable training. To overcome these shortcomings, we introduce lightweight adapters to the pre-trained model and only update them at fine-tuning time. However, existing adapters fail to capture intrinsic temporal relations among video frames or textual words. Moreover, they neglect the preservation of critical task-related information that flows from the raw video-language input into the adapter's low-dimensional space. To address these issues, we first propose a novel REcurrent ADapter (READ) that employs recurrent computation to enable temporal modeling capability. Second, we propose Partial Video-Language Alignment (PVLA) objective via the use of partial optimal transport to maintain task-related information flowing into our READ modules. We validate our READ framework through extensive experiments where READ significantly outperforms all existing fine-tuning strategies on multiple low-resource temporal language grounding and video-language summarization benchmarks. The code, model, and data have been made available at https://nguyentthong.github.io/READ.

Yufei He, Juncheng Liu, Yue Liu et al.

A fundamental limitation of current AI agents is their inability to learn complex skills on the fly at test time, often behaving like "clever but clueless interns" in novel environments. This severely limits their practical utility. To systematically measure and drive progress on this challenge, we first introduce the Jericho Test-Time Learning (J-TTL) benchmark. J-TTL is a new evaluation setup where an agent must play the same game for several consecutive episodes, attempting to improve its performance from one episode to the next. On J-TTL, we find that existing adaptation methods like reflection, memory, or reinforcement learning struggle. To address the challenges posed by our benchmark, we present EvoTest, an evolutionary test-time learning framework that improves an agent without any fine-tuning or gradients-by evolving the entire agentic system after every episode. EvoTest has two roles: the Actor Agent, which plays the game, and the Evolver Agent, which analyzes the episode transcript to propose a revised configuration for the next run. This configuration rewrites the prompt, updates memory by logging effective state-action choices, tunes hyperparameters, and learns the tool-use routines. On our J-TTL benchmark, EvoTest consistently increases performance, outperforming not only reflection and memory-only baselines but also more complex online fine-tuning methods. Notably, our method is the only one capable of winning two games (Detective and Library), while all baselines fail to win any.

Stanislas Ducotterd, Zhiyuan Hu, Michael Unser et al.

Phase retrieval seeks to recover a complex signal from amplitude-only measurements, a challenging nonlinear inverse problem. Current theory and algorithms often ignore signal priors. By contrast, we evaluate here a variety of image priors in the context of severe undersampling with structured random Fourier measurements. Our results show that those priors significantly improve reconstruction, allowing accurate reconstruction even below the weak recovery threshold.

Giuseppe Alessio D'Inverno, Zhiyuan Hu, Leo Davy et al.

Information propagation characterizes how input correlations evolve across layers in deep neural networks. This framework has been well studied using mean-field theory, which assumes infinitely wide networks. However, these assumptions break down for practical, finite-size networks. In this work, we study information propagation in randomly initialized neural networks with finite width and reveal that the boundary between ordered and chaotic regimes exhibits a fractal structure. This shows the fundamental complexity of neural network dynamics, in a setting that is independent of input data and optimization. To extend this analysis beyond multilayer perceptrons, we leverage recently introduced Fourier-based structured transforms, and show that information propagation in convolutional neural networks also follow the same behavior. Our investigation highlights the importance of finite network depth with respect to the tradeoff between separation and robustness.

Yubin Kim, Zhiyuan Hu, Hyewon Jeong et al.

Large Language Models (LLMs) as clinical agents require careful behavioral adaptation. While adept at reactive tasks (e.g., diagnosis reasoning), LLMs often struggle with proactive engagement, like unprompted identification of critical missing information or risks. We introduce BehaviorBench, a comprehensive dataset to evaluate agent behaviors across a clinical assistance spectrum, ranging from reactive query responses to proactive interventions (e.g., clarifying ambiguities, flagging overlooked critical data). Our BehaviorBench experiments reveal LLMs' inconsistent proactivity. To address this, we propose BehaviorSFT, a novel training strategy using behavioral tokens to explicitly condition LLMs for dynamic behavioral selection along this spectrum. BehaviorSFT boosts performance, achieving up to 97.3% overall Macro F1 on BehaviorBench and improving proactive task scores (e.g., from 95.0% to 96.5% for Qwen2.5-7B-Ins). Crucially, blind clinician evaluations confirmed BehaviorSFT-trained agents exhibit more realistic clinical behavior, striking a superior balance between helpful proactivity (e.g., timely, relevant suggestions) and necessary restraint (e.g., avoiding over-intervention) versus standard fine-tuning or explicit instructed agents.

Thong Nguyen, Zhiyuan Hu, Xu Lin et al. · mit

Recent years have witnessed outstanding advances of large vision-language models (LVLMs). In order to tackle video understanding, most of them depend upon their implicit temporal understanding capacity. As such, they have not deciphered important components that contribute to temporal understanding ability, which might limit the potential of these LVLMs for video understanding. In this work, we conduct a thorough empirical study to demystify crucial components that influence the temporal understanding of LVLMs. Our empirical study reveals that significant impacts are centered around the intermediate interface between the visual encoder and the large language model. Building on these insights, we propose a temporal-oriented recipe that encompasses temporal-oriented training schemes and an upscaled interface. Our final model developed using our recipe significantly enhances previous LVLMs on standard video understanding tasks.

Lexiang Tang, Xianwei Zhuang, Bang Yang et al.

Large vision-language models (LVLMs) have demonstrated impressive capabilities across diverse multimodal tasks, yet they remain highly susceptible to visual hallucinations (VH), often producing confident but inaccurate descriptions of visual content. Building on the insight that not all tokens and attention heads contribute equally to VH mitigation, we introduce VisFlow, a lightweight and training-free framework that alleviates hallucinations by directly modulating attention patterns during inference. To address two primary challenges of VH, namely insufficient visual attention and the dominance of language priors, we identify three problematic attention behaviors in LVLMs: (1) disproportionate allocation of attention to uninformative or trailing visual tokens, (2) over-dependence on the previously generated token, and (3) excessive fixation on system prompts that hinders multimodal integration. To overcome these issues, VisFlow introduces a dual-level Attention Intervention, consisting of Token-level Attention Intervention (TAI), which reinforces attention to salient visual regions, and Head-level Attention Intervention (HAI), which suppresses undue focus on system prompts and adjacent text tokens. Together, these interventions strengthen visual alignment while reducing linguistic bias. Extensive experiments across diverse models and benchmarks demonstrate that VisFlow effectively mitigates hallucinations with minimal computational overhead.

Matej Jusup, Kenan Zhang, Zhiyuan Hu et al.

The rapid expansion of ride-sourcing services such as Uber, Lyft, and Didi Chuxing has fundamentally reshaped urban transportation by offering flexible, on-demand mobility via mobile applications. Despite their convenience, these platforms confront significant operational challenges, particularly vehicle rebalancing - the strategic repositioning of a large group of vehicles to address spatiotemporal mismatches in supply and demand. Inadequate rebalancing not only results in prolonged rider waiting times and inefficient vehicle utilization but also leads to fairness issues, such as the inequitable distribution of service quality and disparities in driver income. To tackle these complexities, we introduce continuous-state mean-field control (MFC) and mean-field reinforcement learning (MFRL) models that employ continuous vehicle repositioning actions. MFC and MFRL offer scalable solutions by modeling each vehicle's behavior through interaction with the vehicle distribution, rather than with individual vehicles. This limits the issues arising from the curse of dimensionality inherent in traditional multi-agent methods, enabling coordination across large fleets with significantly reduced computational complexity. To ensure equitable service access across geographic regions, we integrate an accessibility constraint into both models. Extensive empirical evaluation using real-world data-driven simulation of Shenzhen demonstrates the real-time efficiency and robustness of our approach. Remarkably, it scales to tens of thousands of vehicles, with training times comparable to the decision time of a single linear programming rebalancing. Besides, policies generated by our approach effectively explore the efficiency-equity Pareto front, outperforming conventional benchmarks across key metrics like fleet utilization, fulfilled requests, and pickup distance, while ensuring equitable service access.

Fanghua Ye, Zhiyuan Hu, Emine Yilmaz

Dialogue systems have received increasing attention while automatically evaluating their performance remains challenging. User satisfaction estimation (USE) has been proposed as an alternative. It assumes that the performance of a dialogue system can be measured by user satisfaction and uses an estimator to simulate users. The effectiveness of USE depends heavily on the estimator. Existing estimators independently predict user satisfaction at each turn and ignore satisfaction dynamics across turns within a dialogue. In order to fully simulate users, it is crucial to take satisfaction dynamics into account. To fill this gap, we propose a new estimator ASAP (sAtisfaction eStimation via HAwkes Process) that treats user satisfaction across turns as an event sequence and employs a Hawkes process to effectively model the dynamics in this sequence. Experimental results on four benchmark dialogue datasets demonstrate that ASAP can substantially outperform state-of-the-art baseline estimators.

Ziqi Pang, Zhiyuan Hu, Pavel Tokmakov et al.

Virtually all of deep learning literature relies on the assumption of large amounts of available training data. Indeed, even the majority of few-shot learning methods rely on a large set of "base classes" for pretraining. This assumption, however, does not always hold. For some tasks, annotating a large number of classes can be infeasible, and even collecting the images themselves can be a challenge in some scenarios. In this paper, we study this problem and call it "Small Data" setting, in contrast to "Big Data". To unlock the full potential of small data, we propose to augment the models with annotations for other related tasks, thus increasing their generalization abilities. In particular, we use the richly annotated scene parsing dataset ADE20K to construct our realistic Long-tail Recognition with Diverse Supervision (LRDS) benchmark by splitting the object categories into head and tail based on their distribution. Following the standard few-shot learning protocol, we use the head classes for representation learning and the tail classes for evaluation. Moreover, we further subsample the head categories and images to generate two novel settings which we call "Scarce-Class" and "Scarce-Image", respectively corresponding to the shortage of samples for rare classes and training images. Finally, we analyze the effect of applying various additional supervision sources under the proposed settings. Our experiments demonstrate that densely labeling a small set of images can indeed largely remedy the small data constraints.

Haozhe Wu, Zhiyuan Hu, Jia Jia et al.

Online Social Networks (OSNs) evolve through two pervasive behaviors: follow and unfollow, which respectively signify relationship creation and relationship dissolution. Researches on social network evolution mainly focus on the follow behavior, while the unfollow behavior has largely been ignored. Mining unfollow behavior is challenging because user's decision on unfollow is not only affected by the simple combination of user's attributes like informativeness and reciprocity, but also affected by the complex interaction among them. Meanwhile, prior datasets seldom contain sufficient records for inferring such complex interaction. To address these issues, we first construct a large-scale real-world Weibo dataset, which records detailed post content and relationship dynamics of 1.8 million Chinese users. Next, we define user's attributes as two categories: spatial attributes (e.g., social role of user) and temporal attributes (e.g., post content of user). Leveraging the constructed dataset, we systematically study how the interaction effects between user's spatial and temporal attributes contribute to the unfollow behavior. Afterwards, we propose a novel unified model with heterogeneous information (UMHI) for unfollow prediction. Specifically, our UMHI model: 1) captures user's spatial attributes through social network structure; 2) infers user's temporal attributes through user-posted content and unfollow history; and 3) models the interaction between spatial and temporal attributes by the nonlinear MLP layers. Comprehensive evaluations on the constructed dataset demonstrate that the proposed UMHI model outperforms baseline methods by 16.44% on average in terms of precision. In addition, factor analyses verify that both spatial attributes and temporal attributes are essential for mining unfollow behavior.