arXiv subjects

Junlong Li

Publications and source records attributed to Junlong Li.

At least 19 recordsLinked to original sources

arXiv2026

EgoErrorVQA: Assess Egocentric Comprehension Capabilities through Procedural Errors for Ego-Agentic AI

The majority of our everyday activities are procedural and consist of sequences of interdependent steps. However, existing benchmarks for Visual Agents and Visual Language Models (VLMs) overlook the evaluation of their procedural comprehension ability from an egocentric visual perspective, particularly for detecting procedural errors, a critical capability for everyday assistance. To bridge this gap, the EgoErrorVQA task is firstly proposed for egocentric procedural comprehension with explicit procedural errors modeling. Besides, we develop a user-friendly evaluator agent based on the Agent2Agent (A2A) protocol, enabling rigorous and standardized evaluation of visual agents through VQA-based interaction. A range of models are evaluated using both open-ended and multiple-choice questions, revealing persistent weaknesses in handling procedural errors and error types. Moreover, we introduce Ego-ADR, an Adaptive Decoupled Reasoning framework that decouples complex procedural reasoning to enhance models' understanding of procedural errors. It achieves consistent performance gains over the selected baselines and attains state-of-the-art results on several metrics under comparable settings. Code: https://github.com/z1oong/EgoErrorVQA

cs.CV↗

arXiv2026

EgoProceVQA: A Novel Egocentric Procedural Understanding Task with Self-Skill-Exploration Agent

Most daily activities are inherently procedural. However, existing evaluations for egocentric video understanding seldom address procedural understanding and largely overlook complex key-step-level reasoning under the widely used video question answering (VQA) paradigm for MLLMs. Such capabilities are crucial for building procedural AI assistants deployable on wearable devices. To bridge this gap, we introduce the Egocentric Procedural Understanding VQA task (EgoProceVQA), which systematically evaluates egocentric procedural reasoning abilities of current MLLMs and agents through six types of key-step-centric questions. Furthermore, we develop EgoProceGen, a data generation platform that efficiently constructs QA data tailored to different question types. Based on this platform, we build a benchmark with 3,600 questions, four common procedural scenarios, and 31 everyday procedural tasks. Evaluations on EgoProceVQA show that existing MLLMs and agents still have substantial room for improvement in procedural understanding. Therefore, we further propose EgoProceAgent, a self-skill-exploration agentic framework. We design a generic tool library for procedural understanding and a standardized sub-skill library shared across tools and models, enabling self-exploration without ground-truth supervision. By exploring how to compose and select sub-skills, the agent discovers effective skill strategies for diverse problems, and attains state-of-the-art performance among open-source models on multiple tasks. Together, our benchmark, generation platform, and agentic framework establish a unified foundation for EgoProceVQA. Project page: https://z1oong.github.io/EgoProceVQA/.

cs.CV↗

arXiv2026

GUI-C$^2$: Coarse-to-Fine GUI Grounding via Difficulty-Aware Reinforcement Learning

Existing agentic reinforcement learning methods for GUI grounding have limitations at two levels. At the data level, current approaches typically treat all training samples equally, although their training value to the baseline model varies with difficulty. Overlooking this can greatly reduce training efficiency or even cause collapse. At the strategy level, existing frameworks struggle to balance the trade-off between cropping larger regions for sufficient context and smaller ones for reduced redundancy, a tension inherent to tool-augmented grounding agents. In addition, overly complex decision-making is difficult for small-parameter models and significantly increases inference time. To address these issues, at the data level, we propose GUI-D, a data mining and difficulty scoring pipeline that identifies the training-worthy samples by proper testing and assigns difficulty scores to guide subsequent training weights. At the strategy level, we propose GUI-C$^2$, which employs an area-gated coarse-to-fine refinement mechanism that progressively narrows the visual field via model-internal uncertainty signals, adaptively reserving context for large targets while amplifying precision for small ones, reinforced by improvement-aware stage rewards that ensure each refinement genuinely advances grounding. Meanwhile, we simplify the decision-making process to greatly reduce additional inference time. Finally, extensive experiments show that our method achieves state-of-the-art performance. The code and data will be publicly available.

cs.CV↗

arXiv2026

DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence

We present a preview version of DeepSeek-V4 series, including two strong Mixture-of-Experts (MoE) language models -- DeepSeek-V4-Pro with 1.6T parameters (49B activated) and DeepSeek-V4-Flash with 284B parameters (13B activated) -- both supporting a context length of one million tokens. DeepSeek-V4 series incorporate several key upgrades in architecture and optimization: (1) a hybrid attention architecture that combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to improve long-context efficiency; (2) Manifold-Constrained Hyper-Connections (mHC) that enhance conventional residual connections; (3) and the Muon optimizer for faster convergence and greater training stability. We pre-train both models on more than 32T diverse and high-quality tokens, followed by a comprehensive post-training pipeline that unlocks and further enhances their capabilities. DeepSeek-V4-Pro-Max, the maximum reasoning effort mode of DeepSeek-V4-Pro, redefines the state-of-the-art for open models, outperforming its predecessors in core tasks. Meanwhile, DeepSeek-V4 series are highly efficient in long-context scenarios. In the one-million-token context setting, DeepSeek-V4-Pro requires only 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2. This enables us to routinely support one-million-token contexts, thereby making long-horizon tasks and further test-time scaling more feasible. The model checkpoints are available at https://huggingface.co/collections/deepseek-ai/deepseek-v4.

cs.CL↗

arXiv2026

SkillCraft: Can LLM Agents Learn to Use Tools Skillfully?

Real-world tool-using agents operate over long-horizon workflows with recurring structure and diverse demands, where effective behavior requires not only invoking atomic tools but also abstracting, and reusing higher-level tool compositions. However, existing benchmarks mainly measure instance-level success under static tool sets, offering limited insight into agents' ability to acquire such reusable skills. We address this gap by introducing SkillCraft, a benchmark explicitly stress-test agent ability to form and reuse higher-level tool compositions, where we call Skills. SkillCraft features realistic, highly compositional tool-use scenarios with difficulty scaled along both quantitative and structural dimensions, designed to elicit skill abstraction and cross-task reuse. We further propose a lightweight evaluation protocol that enables agents to auto-compose atomic tools into executable Skills, cache and reuse them inside and across tasks, thereby improving efficiency while accumulating a persistent library of reusable skills. Evaluating state-of-the-art agents on SkillCraft, we observe substantial efficiency gains, with token usage reduced by up to 80% by skill saving and reuse. Moreover, success rate strongly correlates with tool composition ability at test time, underscoring compositional skill acquisition as a core capability.

cs.CL↗

arXiv2025

DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models

We introduce DeepSeek-V3.2, a model that harmonizes high computational efficiency with superior reasoning and agent performance. The key technical breakthroughs of DeepSeek-V3.2 are as follows: (1) DeepSeek Sparse Attention (DSA): We introduce DSA, an efficient attention mechanism that substantially reduces computational complexity while preserving model performance in long-context scenarios. (2) Scalable Reinforcement Learning Framework: By implementing a robust reinforcement learning protocol and scaling post-training compute, DeepSeek-V3.2 performs comparably to GPT-5. Notably, our high-compute variant, DeepSeek-V3.2-Speciale, surpasses GPT-5 and exhibits reasoning proficiency on par with Gemini-3.0-Pro, achieving gold-medal performance in both the 2025 International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI). (3) Large-Scale Agentic Task Synthesis Pipeline: To integrate reasoning into tool-use scenarios, we developed a novel synthesis pipeline that systematically generates training data at scale. This methodology facilitates scalable agentic post-training, yielding substantial improvements in generalization and instruction-following robustness within complex, interactive environments.

cs.CL↗

arXiv2025

Building Egocentric Procedural AI Assistant: Methods, Benchmarks, and Challenges

Driven by recent advances in vision-language models (VLMs) and egocentric perception research, the emerging topic of an egocentric procedural AI assistant (EgoProceAssist) is introduced to step-by-step support daily procedural tasks in a first-person view. In this paper, we start by identifying three core tasks in EgoProceAssist: egocentric procedural error detection, egocentric procedural learning, and egocentric procedural question answering, then introduce two enabling dimensions: real-time and streaming video understanding, and proactive interaction in procedural contexts. We define these tasks within a new taxonomy as the EgoProceAssist's essential functions and illustrate how they can be deployed in real-world scenarios for daily activity assistants. Specifically, our work encompasses a comprehensive review of current techniques, relevant datasets, and evaluation metrics across these five core areas. To clarify the gap between the proposed EgoProceAssist and existing VLM-based assistants, we conduct novel experiments to provide a comprehensive evaluation of representative VLM-based methods. Through these findings and our technical analysis, we discuss the challenges ahead and suggest future research directions. Furthermore, an exhaustive list of this study is publicly available in an active repository that continuously collects the latest work: https://github.com/z1oong/Building-Egocentric-Procedural-AI-Assistant.

cs.CV↗

arXiv2025

The Tool Decathlon: Benchmarking Language Agents for Diverse, Realistic, and Long-Horizon Task Execution

Real-world language agents must handle complex, multi-step workflows across diverse Apps. For instance, an agent may manage emails by coordinating with calendars and file systems, or monitor a production database to detect anomalies and generate reports following an operating manual. However, existing language agent benchmarks often focus on narrow domains or simplified tasks that lack the diversity, realism, and long-horizon complexity required to evaluate agents' real-world performance. To address this gap, we introduce the Tool Decathlon (dubbed as Toolathlon), a benchmark for language agents offering diverse Apps and tools, realistic environment setup, and reliable execution-based evaluation. Toolathlon spans 32 software applications and 604 tools, ranging from everyday platforms such as Google Calendar and Notion to professional ones like WooCommerce, Kubernetes, and BigQuery. Most of the tools are based on a high-quality set of Model Context Protocol (MCP) servers that we may have revised or implemented ourselves. Unlike prior works, which primarily ensure functional realism but offer limited environment state diversity, we provide realistic initial environment states from real software, such as Canvas courses with dozens of students or real financial spreadsheets. This benchmark includes 108 manually sourced or crafted tasks in total, requiring interacting with multiple Apps over around 20 turns on average to complete. Each task is strictly verifiable through dedicated evaluation scripts. Comprehensive evaluation of SOTA models highlights their significant shortcomings: the best-performing model, Claude-4.5-Sonnet, achieves only a 38.6% success rate with 20.2 tool calling turns on average, while the top open-weights model DeepSeek-V3.2-Exp reaches 20.1%. We expect Toolathlon to drive the development of more capable language agents for real-world, long-horizon task execution.

cs.CL↗

arXiv2025

CodeI/O: Condensing Reasoning Patterns via Code Input-Output Prediction

Reasoning is a fundamental capability of Large Language Models. While prior research predominantly focuses on enhancing narrow skills like math or code generation, improving performance on many other reasoning tasks remains challenging due to sparse and fragmented training data. To address this issue, we propose CodeI/O, a novel approach that systematically condenses diverse reasoning patterns inherently embedded in contextually-grounded codes, through transforming the original code into a code input-output prediction format. By training models to predict inputs/outputs given code and test cases entirely in natural language as Chain-of-Thought (CoT) rationales, we expose them to universal reasoning primitives -- like logic flow planning, state-space searching, decision tree traversal, and modular decomposition -- while decoupling structured reasoning from code-specific syntax and preserving procedural rigor. Experimental results demonstrate CodeI/O leads to consistent improvements across symbolic, scientific, logic, math & numerical, and commonsense reasoning tasks. By matching the existing ground-truth outputs or re-executing the code with predicted inputs, we can verify each prediction and further enhance the CoTs through multi-turn revision, resulting in CodeI/O++ and achieving higher performance. Our data and models are available at https://github.com/hkust-nlp/CodeIO.

cs.CL↗

arXiv2025

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

General reasoning represents a long-standing and formidable challenge in artificial intelligence. Recent breakthroughs, exemplified by large language models (LLMs) and chain-of-thought prompting, have achieved considerable success on foundational reasoning tasks. However, this success is heavily contingent upon extensive human-annotated demonstrations, and models' capabilities are still insufficient for more complex problems. Here we show that the reasoning abilities of LLMs can be incentivized through pure reinforcement learning (RL), obviating the need for human-labeled reasoning trajectories. The proposed RL framework facilitates the emergent development of advanced reasoning patterns, such as self-reflection, verification, and dynamic strategy adaptation. Consequently, the trained model achieves superior performance on verifiable tasks such as mathematics, coding competitions, and STEM fields, surpassing its counterparts trained via conventional supervised learning on human demonstrations. Moreover, the emergent reasoning patterns exhibited by these large-scale models can be systematically harnessed to guide and enhance the reasoning capabilities of smaller models.

cs.CL↗

arXiv2024

DeepSeek-V3 Technical Report

We present DeepSeek-V3, a strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token. To achieve efficient inference and cost-effective training, DeepSeek-V3 adopts Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were thoroughly validated in DeepSeek-V2. Furthermore, DeepSeek-V3 pioneers an auxiliary-loss-free strategy for load balancing and sets a multi-token prediction training objective for stronger performance. We pre-train DeepSeek-V3 on 14.8 trillion diverse and high-quality tokens, followed by Supervised Fine-Tuning and Reinforcement Learning stages to fully harness its capabilities. Comprehensive evaluations reveal that DeepSeek-V3 outperforms other open-source models and achieves performance comparable to leading closed-source models. Despite its excellent performance, DeepSeek-V3 requires only 2.788M H800 GPU hours for its full training. In addition, its training process is remarkably stable. Throughout the entire training process, we did not experience any irrecoverable loss spikes or perform any rollbacks. The model checkpoints are available at https://github.com/deepseek-ai/DeepSeek-V3.

cs.CL↗

arXiv2024

Diving into Self-Evolving Training for Multimodal Reasoning

Self-evolving trainin--where models iteratively learn from their own outputs--has emerged as a key approach for complex reasoning tasks, addressing the scarcity of high-quality chain-of-thought data. However, its effectiveness in multimodal reasoning, a domain more intricate than text-only reasoning, remains underexplored, and the understanding of critical factors in this training paradigm remains limited. Furthermore, a central challenge for this training method is performance saturation, which impedes further improvements and scalability. Inspired by reinforcement learning (RL), in this paper, we reframe self-evolving training for multimodal reasoning through the lens of RL, identifying three pivotal factors: Training Method, Reward Model, and Prompt Variation. Through systematic analysis, we establish relatively optimal design principles that significantly enhance multimodal reasoning capabilities. Moreover, delving deeper into training dynamics, we uncover the roots of saturation and propose a new automatic balancing mechanism to mitigate this limitation. Building on these insights, we propose M-STAR (Multimodal Self-evolving Training for Reasoning), a framework that achieves consistent performance gains across models of varying sizes and diverse benchmarks. All resources are made publicly available at https://mstar-lmm.github.io.

cs.CL↗

arXiv2024

Programming Every Example: Lifting Pre-training Data Quality Like Experts at Scale

Large language model pre-training has traditionally relied on human experts to craft heuristics for improving the corpora quality, resulting in numerous rules developed to date. However, these rules lack the flexibility to address the unique characteristics of individual example effectively. Meanwhile, applying tailored rules to every example is impractical for human experts. In this paper, we demonstrate that even small language models, with as few as 0.3B parameters, can exhibit substantial data refining capabilities comparable to those of human experts. We introduce Programming Every Example (ProX), a novel framework that treats data refinement as a programming task, enabling models to refine corpora by generating and executing fine-grained operations, such as string normalization, for each individual example at scale. Experimental results show that models pre-trained on ProX-curated data outperform either original data or data filtered by other selection methods by more than 2% across various downstream benchmarks. Its effectiveness spans various model sizes and pre-training corpora, including C4, RedPajama-V2, FineWeb, FineWeb-Edu, and DCLM. Furthermore, ProX exhibits significant potential in domain-specific continual pre-training: without domain specific design, models trained on OpenWebMath refined by ProX outperform human-crafted rule-based methods, improving average accuracy by 7.6% over Mistral-7B, with 14.6% for Llama-2-7B and 20.3% for CodeLlama-7B, all within 10B tokens to be comparable to models like Llemma-7B trained on 200B tokens. Further analysis highlights that ProX significantly saves training FLOPs, offering a promising path for efficient LLM pre-training. We are open-sourcing ProX with >500B corpus, models, and sharing all training and implementation details for reproducible research and future innovation. Code: https://github.com/GAIR-NLP/ProX

cs.CL↗

arXiv2024

Reformatted Alignment

The quality of finetuning data is crucial for aligning large language models (LLMs) with human values. Current methods to improve data quality are either labor-intensive or prone to factual errors caused by LLM hallucinations. This paper explores elevating the quality of existing instruction data to better align with human values, introducing a simple and effective approach named ReAlign, which reformats the responses of instruction data into a format that better aligns with pre-established criteria and the collated evidence. This approach minimizes human annotation, hallucination, and the difficulty in scaling, remaining orthogonal to existing alignment techniques. Experimentally, ReAlign significantly boosts the general alignment ability, math reasoning, factuality, and readability of the LLMs. Encouragingly, without introducing any additional data or advanced training techniques, and merely by reformatting the response, LLaMA-2-13B's mathematical reasoning ability on GSM8K can be improved from 46.77% to 56.63% in accuracy. Additionally, a mere 5% of ReAlign data yields a 67% boost in general alignment ability measured by the Alpaca dataset. This work highlights the need for further research into the science and mechanistic interpretability of LLMs. We have made the associated code and data publicly accessible to support future studies at https://github.com/GAIR-NLP/ReAlign.

cs.CL↗

arXiv2024

Dissecting Human and LLM Preferences

As a relative quality comparison of model responses, human and Large Language Model (LLM) preferences serve as common alignment goals in model fine-tuning and criteria in evaluation. Yet, these preferences merely reflect broad tendencies, resulting in less explainable and controllable models with potential safety risks. In this work, we dissect the preferences of human and 32 different LLMs to understand their quantitative composition, using annotations from real-world user-model conversations for a fine-grained, scenario-wise analysis. We find that humans are less sensitive to errors, favor responses that support their stances, and show clear dislike when models admit their limits. On the contrary, advanced LLMs like GPT-4-Turbo emphasize correctness, clarity, and harmlessness more. Additionally, LLMs of similar sizes tend to exhibit similar preferences, regardless of their training methods, and fine-tuning for alignment does not significantly alter the preferences of pretrained-only LLMs. Finally, we show that preference-based evaluation can be intentionally manipulated. In both training-free and training-based settings, aligning a model with the preferences of judges boosts scores, while injecting the least preferred properties lowers them. This results in notable score shifts: up to 0.59 on MT-Bench (1-10 scale) and 31.94 on AlpacaEval 2.0 (0-100 scale), highlighting the significant impact of this strategic adaptation. Interactive Demo: https://huggingface.co/spaces/GAIR/Preference-Dissection-Visualization Dataset: https://huggingface.co/datasets/GAIR/preference-dissection Code: https://github.com/GAIR-NLP/Preference-Dissection

cs.CL↗

arXiv2024

Extending LLMs' Context Window with 100 Samples

Large Language Models (LLMs) are known to have limited extrapolation ability beyond their pre-trained context window, constraining their application in downstream tasks with lengthy inputs. Recent studies have sought to extend LLMs' context window by modifying rotary position embedding (RoPE), a popular position encoding method adopted by well-known LLMs such as LLaMA, PaLM, and GPT-NeoX. However, prior works like Position Interpolation (PI) and YaRN are resource-intensive and lack comparative experiments to assess their applicability. In this work, we identify the inherent need for LLMs' attention entropy (i.e. the information entropy of attention scores) to maintain stability and introduce a novel extension to RoPE which combines adjusting RoPE's base frequency and scaling the attention logits to help LLMs efficiently adapt to a larger context window. We validate the superiority of our method in both fine-tuning performance and robustness across different context window sizes on various context-demanding tasks. Notably, our method extends the context window of LLaMA-2-7B-Chat to 16,384 with only 100 samples and 6 training steps, showcasing extraordinary efficiency. Finally, we also explore how data compositions and training curricula affect context window extension for specific downstream tasks, suggesting fine-tuning LLMs with lengthy conversations as a good starting point. We release our code and SFT data at https://github.com/GAIR-NLP/Entropy-ABF.

cs.CL↗

arXiv2024

The Critique of Critique

Critique, as a natural language description for assessing the quality of model-generated content, has played a vital role in the training, evaluation, and refinement of LLMs. However, a systematic method to evaluate the quality of critique is lacking. In this paper, we pioneer the critique of critique, termed MetaCritique, which builds specific quantification criteria. To achieve a reliable evaluation outcome, we propose Atomic Information Units (AIUs), which describe the critique in a more fine-grained manner. MetaCritique aggregates each AIU's judgment for the overall score. Moreover, MetaCritique delivers a natural language rationale for the intricate reasoning within each judgment. Lastly, we construct a meta-evaluation dataset covering 4 tasks across 16 public datasets involving human-written and LLM-generated critiques. Experiments demonstrate that MetaCritique can achieve near-human performance. Our study can facilitate future research in LLM critiques based on our following observations and released resources: (1) superior critiques judged by MetaCritique can lead to better refinements, indicating that it can potentially enhance the alignment of existing LLMs; (2) the leaderboard of critique models reveals that open-source critique models commonly suffer from factuality issues; (3) relevant code and data are publicly available at https://github.com/GAIR-NLP/MetaCritique to support deeper exploration; (4) an API at PyPI with the usage documentation in Appendix C allows users to assess the critique conveniently.

cs.CL↗

arXiv2023

Self-prompted Chain-of-Thought on Large Language Models for Open-domain Multi-hop Reasoning

In open-domain question-answering (ODQA), most existing questions require single-hop reasoning on commonsense. To further extend this task, we officially introduce open-domain multi-hop reasoning (ODMR) by answering multi-hop questions with explicit reasoning steps in open-domain setting. Recently, large language models (LLMs) have found significant utility in facilitating ODQA without external corpus. Furthermore, chain-of-thought (CoT) prompting boosts the reasoning capability of LLMs to a greater extent with manual or automated paradigms. However, existing automated methods lack of quality assurance, while manual approaches suffer from limited scalability and poor diversity, hindering the capabilities of LLMs. In this paper, we propose Self-prompted Chain-of-Thought (SP-CoT), an automated framework to mass-produce high quality CoTs of LLMs, by LLMs and for LLMs. SP-CoT introduces an automated generation pipeline of high quality ODMR datasets, an adaptive sampler for in-context CoT selection and self-prompted inference via in-context learning. Extensive experiments on four multi-hop question-answering benchmarks show that our proposed SP-CoT not only significantly surpasses the previous SOTA methods on large-scale (175B) LLMs, but also nearly doubles the zero-shot performance of small-scale (13B) LLMs. Further analysis reveals the remarkable capability of SP-CoT to elicit direct and concise intermediate reasoning steps by recalling $\sim$50\% of intermediate answers on MuSiQue-Ans dataset.

cs.CL↗