arXiv subjects

Jia Li

Publications and source records attributed to Jia Li.

At least 91 records · Page 5Linked to original sources

arXiv2026

CoACT: Action-Preserving Observation Compression for Coding Agents

LLM-based coding agents solve software-engineering tasks through iterative interactions with development environments, where returned observations accumulate in the context and become a major source of inference cost. Observation compression reduces this cost by shortening observations before they are appended to the context. However, existing methods still exhibit an unsatisfactory efficiency-effectiveness trade-off, as they do not explicitly model how compression affects the agent's subsequent behavior. This paper proposes CoACT, an action-preserving observation compression method for coding agents. CoACT is built on next-action preservation (NAP), which requires a compressed observation to induce the same next action as the raw observation. By checking the agent's immediate next action, NAP provides a practical signal for whether a compression preserves the information needed for continued task solving. During training, a teacher model first generates multiple compressed candidates of each observation. CoACT then uses an action-preservation reward based on NAP to filter out candidates that would change the agent's next action, and uses a length-reduction reward to choose compact candidates as supervision for a lightweight compressor. Experiments on SWE-bench Verified with three agentic models show that CoACT reduces average total token consumption by 33.0% while maintaining task-solving effectiveness close to the uncompressed agent.

cs.SE↗

arXiv2026

RealBench: A Repo-Level Code Generation Benchmark Aligned with Real-World Software Development Practices

Writing code requires significant time and effort in software development. To automate this process, researchers have made substantial progress using Large Language Models (LLMs) for code generation. Many benchmarks like HumanEval and EvoCodeBench have been created to evaluate LLMs by requiring them to generate code from natural language requirements. However, in enterprise applications and team development, developers typically write code based on structured designs or specifications rather than raw natural language descriptions. This gap between existing benchmarks and real industry development practices means that current benchmark scores may not accurately reflect how much code generation can help automate software development tasks. To address this gap, we propose RealBench, a repository-level code generation benchmark aligned with real-world industry software development practices. Each example includes both natural language requirements and UML diagrams as system design, matching how developers typically receive specifications. Based on the constructed benchmarks, we conduct a systematic evaluation of advanced LLMs' code generation capabilities when provided with structured system designs. The experimental results reveal key insights in current LLMs' capabilities for repo-level code generation aligned with real-world software development practices. First, we notice that regarding repo-level code generation, LLMs show much worse performance and there are significant performance gaps among LLMs. Second, LLMs are good at finding and creating modules defined in UML diagrams, but the quality of generated modules is often poor due to grammar and logic errors. Third, generating the entire repository at once is the best generation strategy on smaller repositories, while generating a complex repository with the module-by-module strategy works better compared to other strategies.

cs.SE↗

arXiv2026

Towards Memory-Efficient Autoregressive Video Generation via Instance-Specific Parametric Absorption

Autoregressive (AR) streaming models have emerged as a powerful paradigm for long video generation. However, the linearly growing Key-Value (KV) cache poses a significant bottleneck, leading to memory overload and degraded inference throughput. A common compression method is to drop redundant KV tokens, which often breaks long-range dependencies, resulting in temporal flickering and identity loss. In this paper, we propose Instance-Specific Parametric Absorption (ISPA), a novel framework that shifts the KV cache compression from discarding to distilling. The core idea is to transit a subset of layers from Full-Attention (F-Layers) to memory-efficient Local-Attention (L-Layers) by "absorbing" historical context into the model's weights. Specifically, during a brief warmup phase, ISPA monitors the output discrepancy between global and local attention. At the transition point, we solve a closed-form least-squares problem to compute an instance-specific weight modulation that compensates for the missing history. Experiments across architectures (1.3B to 14B) demonstrate that ISPA can remove up to 50\% of the KV cache with near-lossless visual quality. We hope this perspective encourages future work to explore parametric memory consolidation beyond external token-level cache management for streaming generative models.

cs.CV↗

arXiv2026

How Much Future Helps? A Controlled Study of Future-Privileged Supervision for Causal Egocentric Gaze Estimation

Egocentric gaze estimation is commonly studied using models that process the full video with access to future frames, while real-world applications require strictly causal, online prediction. This discrepancy raises key questions: Does future context inherently provide valuable signals for gaze estimation? If so, how much future look-ahead optimally supervises a causal model during training? To investigate, we propose a controlled framework featuring a future-aware branch that accesses a tunable look-ahead horizon during training but is discarded at inference. This design isolates the impact of future context while keeping the inference architecture fixed and strictly causal. Across EGTEA Gaze+ and Ego4D, we find that future-privileged supervision consistently improves causal gaze prediction, confirming its utility. However, performance gains do not increase monotonically with longer look-ahead, but rather peak within a bounded temporal regime. Specifically, optimal performance corresponds to roughly 1.7--3.3 seconds of future context ($H{\in}[5, 10]$) on EGTEA Gaze+ and 2.7 seconds ($H{=}10$) on Ego4D. Our results demonstrate that lightweight causal models can effectively absorb future-aware signals, providing practical guidance for real-time egocentric gaze modeling.

cs.CV↗

arXiv2026

VulKey: Automated Vulnerability Repair Guided by Domain-Specific Repair Patterns

The increasing prevalence of software vulnerabilities highlights the need for effective Automatic Vulnerability Repair (AVR) tools. While LLM-based approaches are promising, they struggle to incorporate structured security knowledge from sources like CWE and NVD. Current methods either use this information superficially by concatenating the CWE-ID into the input prompt, yielding negligible benefits, or rely on few-shot learning with rigid, non-generalizable examples, which limits their effectiveness in real-world scenarios. To address this gap, we propose VulKey, an LLM-based AVR framework that leverages a hierarchical abstraction of expert knowledge to guide patch generation. Our novel three-level abstraction formulates repair strategies in terms of CWE type, syntactic actions, and semantic key elements. This approach captures the essence of a security fix with greater generality than concrete examples and more semantic richness than traditional syntax-based templates, overcoming the coverage limitations of prior methods. VulKey is implemented as a two-stage pipeline: first, expert knowledge matching predicts an appropriate repair pattern for the vulnerability; second, repair code generation uses a pattern-guided, fine-tuned LLM to produce secure patches. On the real-world C/C++ dataset PrimeVul, VulKey achieves 31.5% repair accuracy, surpassing the best baseline by 7.6% and outperforming leading tools such as VulMaster and GPT-5. Moreover, VulKey demonstrates cross-language and cross-model generalizability, with state-of-the-art performance on the Java benchmark Vul4J. These results underscore the importance of structured expert knowledge in advancing AVR effectiveness. Our work demonstrates that explicitly modeling and integrating expert security knowledge through hierarchical patterns is a crucial step toward building more effective and reliable AVR tools.

cs.CR↗

arXiv2026

Capture Timing-Attention of Events in Clinical Time Series

The contemporary paradigm of trajectory learning operates fundamentally at the level of group dynamics, systematically reducing individual-level complexity to fit group-level models, thus rendering effective patient subtyping difficult and individual-level modeling largely out of reach. We propose a data-driven paradigm that introduces a dedicated individual-level temporal variable to capture \emph{Timing Attention} (i.e., the degree of concentration of an event's timing distribution across the patient cohort), thereby rendering timing a \emph{computable dimension} that enables individualized temporal features in trajectory learning. Instantiated as the Level-of-Individual Time Transformation (LITT) and applied to longitudinal EHR data from 3,276 breast cancer patients, the proposed paradigm demonstrates, for the first time to our knowledge: (1) automatic discovery of clinically significant patient trajectories, and (2) counterfactual timing deduction, that is, a \emph{What-If Machine}. Both results are purely data-driven, requiring no prior domain knowledge. LITT further achieves strong performance on timing prediction and survival analysis tasks.

cs.LG↗

arXiv2026

Holmes: Multimodal Agentic Diagnosis for Mixed-Language Mobile Crashes at Industrial Scale

Diagnosing mobile crashes in ultra-large-scale industrial applications is a formidable challenge due to the sheer volume of code, the complexity of mixed-language environments, and the inability to reproduce failures locally. Traditional static analysis struggles with scalability, while existing LLM-based agents often rely on reproducible environments unavailable in post-mortem scenarios. We present Holmes, a multi-agent system that automates root cause analysis by synthesizing multimodal runtime signals--stack traces, logs, and thread states--to reconstruct failure contexts without reproduction. Holmes introduces a hierarchical Retrieve-Explore-Reason architecture that leverages low-level artifacts (e.g., registers, assembly) to bridge the semantic gap between open-source business logic and closed-source system frameworks. By dynamically compressing the search space using runtime clues, Holmes precisely navigates 70-million-line codebases to identify non-local defects. Evaluated on real-world crashes from WeChat, Holmes achieves 87.6% accuracy in function-level fault localization and reduces average investigation time by over 98% (to ~77 seconds), demonstrating its effectiveness in transforming labor-intensive debugging into an efficient verification workflow.

cs.AI↗

arXiv2026

Canonical Variates in Wasserstein Metric Space

In this paper, we address the classification of instances represented by distributions on a vector space rather than single points. We consider classification algorithms based on pairwise distances, specifically, the Wasserstein metric between distributions. Central to our investigation is dimension reduction within the Wasserstein metric space to enhance classification accuracy. We introduce a novel approach grounded in the principle of maximizing Fisher's ratio, defined as the quotient of between-class variation to within-class variation. The directions in which this ratio is maximized are termed discriminant coordinates or canonical variates axes. In practice, both between-class and within-class variations are defined as the average squared Wasserstein distances between pairs of distributions, with the pairs either belonging to the same class or to different classes. This ratio optimization is achieved through an iterative algorithm, which alternates between optimal transport and maximization steps within the vector space. Empirical studies are conducted to assess the algorithm's convergence; and experimental results demonstrate that the dimension reduction technique substantially enhances classification performance. Moreover, the new method outperforms well-established algorithms that operate on vector representations derived from distributional data. It also exhibits robustness to variations in how instances are summarized by distributions, such as the number of components in a Gaussian mixture model (GMM) representation.

stat.ML↗

arXiv2026

EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective

Recent benchmarks for Large Language Model (LLM) agents mainly evaluate reasoning, planning, and execution. However, memory is also essential for agents, as it enables them to store, update, and retrieve information over time. This ability remains under-evaluated, largely because existing benchmarks do not provide a systematic way to assess memory mechanisms. In this paper, we study agent memory from a self-evolving perspective and introduce EvoMemBench, a unified benchmark organized along two axes: memory scope (in-episode vs. cross-episode) and memory content (knowledge-oriented vs. execution-oriented). We compare 15 representative memory methods with strong long-context baselines under a standardized protocol. Results show that current memory systems are still far from a general solution: long-context baselines remain highly competitive, memory helps most when the current context is insufficient or tasks are difficult, and no single memory form works consistently across all settings. Retrieval-based methods remain strong for knowledge-intensive settings, whereas procedural and long-term memory methods are more effective for execution-oriented tasks when their stored experience matches the task structure. We hope EvoMemBench facilitates future research on more effective memory systems for LLM-based agents. Our code is available at https://github.com/DSAIL-Memory/EvoMemBench.

cs.CL↗

arXiv2026

Ling and Ring 2.6 Technical Report: Efficient and Instant Agentic Intelligence at Trillion-Parameter Scale

Efficient and scalable agentic intelligence requires models that can deliver both low-latency responses and strong reasoning capabilities while remaining practical to train, serve, and deploy. In this report, we present Ling-2.6 and Ring-2.6, a family of models designed to address this challenge at scale. Ling-2.6 is optimized for instant response generation and high capability per output token, whereas Ring-2.6 is tailored for deeper reasoning and more advanced agentic workflows. Instead of training from scratch, we upgrade the Ling-2.0 base model through architectural migration pre-training and large-scale post-training. This upgrade is guided by a unified co-design of model architecture, optimization objectives, serving systems, and agent training environments, enabling improvements in both model capability and deployment efficiency. At the architectural level, we introduce a hybrid linear attention design that integrates Lightning Attention with MLA, improving the efficiency of long-context training and decoding. To further enhance token efficiency, we optimize capability per output token through Evolutionary Chain-of-Thought, Linguistic Unit Policy Optimization, bidirectional preference alignment, and shortest-correct-response distillation. For agentic capabilities, we propose KPop, a reinforcement learning framework designed to support stable training of Ring-2.6-1T on large-scale environment-grounded data. KPop improves training efficiency through asynchronous scheduling across coding, search, tool use, and workflow execution, enabling scalable learning from complex agent-environment interactions. Together, Ling-2.6 and Ring-2.6 provide a practical pathway toward efficient, scalable, and open agentic systems. We open-source all checkpoints in the 2.6 family to support further research and development in practical agentic intelligence.

cs.CL↗

arXiv2026

Adapting Vision-Language Models from Iconic to Inclusive for Multi-Label Recognition Without Labels

Understanding multi-label images remains a challenging task in computer vision. With the rapid progress of vision-language multimodal learning, vision-language models (VLMs) enable zero-shot recognition without labeled data. However, due to their intrinsic design, these models often prioritize the most iconic object and omit other contextual positives. This intrinsic bias conflicts with the nature of multi-label learning, thereby limiting their applicability. In this work, we propose an unsupervised framework that adapts VLMs from iconic recognition toward inclusive understanding, enabling label-free multi-label image recognition. Our approach consists of two key stages, ``cutting'' and ``sewing'': In the cutting stage, we present the multi-sampling response estimator to prevent the model from concentrating only on one single object. In the second sewing stage, the multi-object blend adaptation is introduced to adjust the labels to better conform to the multi-label distribution while preserving the intrinsic characteristics of the original model within only one epoch. Extensive experiments show that our framework significantly outperforms existing unsupervised approaches on four public datasets, even surpassing several representative weakly supervised baselines. These results demonstrate the potential of adapting pre-trained VLMs for more comprehensive visual understanding without manual annotations. Our code is publicly available at https://github.com/iCVTEAM/TailorCLIP.

cs.CV↗

arXiv2026

Grammar-Constrained Decoding Can Jailbreak LLMs into Generating Malicious Code

Large Language Models (LLMs) are increasingly used for code generation, raising concerns that they may be misused to produce malicious code. Meanwhile, Grammar-Constrained Decoding (GCD) has been widely adopted to improve the reliability of LLM-generated code by enforcing syntactic validity. In this paper, we reveal a counterintuitive risk: this reliability-oriented technique can itself become an attack surface. We uncover a new jailbreak attack, termed CodeSpear, that exploits GCD to induce LLMs into generating malicious code. Our experiments show that simply applying a benign code grammar constraint can effectively jailbreak LLMs. To address this vulnerability, we propose CodeShield, a safety alignment approach that robustly preserves safe behavior even under attacker-controlled grammar constraints. CodeShield aligns the model in the code modality by teaching it to generate honeypot code under GCD. Such code is semantically harmless, so it does not implement the malicious request, and structurally diverse, so it is difficult to suppress through grammar tightening. At the same time, CodeShield still preserves natural-language refusals when natural language is available. Experiments on 10 popular LLMs across 4 benchmarks show that CodeSpear outperforms representative jailbreak baselines and increases the attack success rate by more than 30 percentage points on average. CodeShield also restores safety under CodeSpear while preserving benign utility. Our findings reveal a fundamental risk of GCD and call for greater attention to its potential security implications.

cs.CR↗

arXiv2026

Missing-Token Prompted Reliability-Aware Fusion for Robust Polyglot Speaker Identification

Accurate and robust multimodal speaker identification is essential for multimedia understanding and biometric authentication. However, real-world polyglot scenarios pose two key challenges: speaker-discriminative representations should generalize across languages, and the model should remain reliable when face information is unavailable. To address these challenges, we propose MRAF, a Missing-Token Prompted Reliability-Aware Fusion framework for polyglot speaker identification across complete-modality, missing-face, and cross-lingual scenarios. MRAF represents unavailable face inputs with a learnable missing token instead of fixed zero-valued features, providing a trainable representation of the missing visual state. This design reduces the distribution gap caused by missing inputs and allows subsequent reliability estimation and cross-modal fusion to operate within a unified token space. To adaptively integrate modalities with different reliability, MRAF further introduces a reliability-aware cross-attention fusion module, which estimates face and audio reliability scores, normalizes them into modality weights, and applies these weights to token representations before bidirectional cross-attention. In this way, the model can emphasize reliable modality cues while suppressing unreliable ones. During training, MRAF jointly optimizes multi-branch classification losses, audio-only knowledge distillation, and center loss to improve speaker discrimination and missing-modality robustness. Experiments on the official POLY-SIM 2026 test set demonstrate the effectiveness of the proposed framework. In the final evaluation, MRAF achieves 100% accuracy on P3 and P5, and obtains competitive results on the more challenging missing-face settings P4 and P6. The source code will be released at https://github.com/MSA-LMC/MRAF.

cs.SD↗

arXiv2026

Efficient Scaling of LLM Training with Flexible Context Parallelism

Scaling long-context capabilities is crucial for Large Language Models (LLMs). However, real-world data contain a large number of sequences with heterogeneous lengths. Existing training libraries for LLMs rely on static parallelism strategies, which suffer from severe load imbalance, redundant communication, and suboptimal hardware utilization under data heterogeneity. In this work, we propose Flexible Context Parallelism (FCP), an efficient parallelism strategy that adaptively reconfigures communication groups and context parallelism degrees during LLM training. We generalize more flexible non-power-of-two parallelism degrees and develop a polynomial-time algorithm to generate near-optimal parallelism strategies with only millisecond-level overhead per training batch. FCP is able to maintain high hardware efficiency even under extreme data heterogeneity. Experimental results demonstrate that FCP significantly outperforms Megatron-LM and DeepSpeed in both LLM and MLLM training, achieving up to 1.46x speedup in average throughput while maintaining near-linear scaling efficiency across large-scale clusters. For extremely unbalanced batches, FCP even achieves 2.24x speedup.

cs.DC↗

arXiv2026

Breaking the Tokenizer Barrier: On-Policy Distillation across Model Families

On-Policy Distillation (OPD) has become a core technique in the post-training of Large Language Models (LLMs) for transferring knowledge from domain experts to student models. However, existing OPD distillation methods require teacher and student models to share the same tokenizer, restricting the applicability of OPD within the model series. Current mainstream practice typically employs Supervised Fine-Tuning (SFT) on teacher-generated responses for cross-tokenizer distillation, which fails to capture the rich knowledge embedded in the teacher's probability distribution. In this work, we enable the standard on-policy distillation method to operate across model families, ensuring that high-fidelity token-level signals can propagate across different tokenizers with a precise token-mapping algorithm. Extensive experiments show that cross-tokenizer OPD is significantly more compute-efficient than baselines on various benchmarks. Our results unlock a broader range of teacher-student pairs for OPD, opening up new avenues for adapting and enhancing interactions between LLMs.

cs.LG↗

arXiv2026

Chiseling Out Efficiency: Structured Skeleton Supervision for Efficient Code Generation

Large Language Models (LLMs) are capable of generating syntactically correct and functionally complete programs, greatly streamlining software development. However, recent studies reveal that these programs typically execute substantially slower than human-optimized counterparts. Existing approaches to bridging this efficiency gap typically involve either iteratively optimizing code after generation or fine-tuning models on corpora of efficient code. Yet, these methods expose the model to efficiency signals only by mimicking complete, optimized solutions, without explicitly encoding the structural code patterns essential for achieving high runtime performance. Addressing this gap presents two core challenges: (1) extracting and representing latent, efficiency-oriented structural patterns embedded within complex syntax and control flows, and (2) effectively learning these patterns without destabilizing the semantic training of LLMs. To tackle these challenges, we propose EffiSkel, an efficiency skeleton-guided framework that explicitly extracts and learns efficiency skeletons-abstract, reusable structural patterns underpinning efficient code-by leveraging three complementary strategies. These skeletons are integrated into a multi-task learning regime that jointly optimizes code generation and skeleton prediction. Experiments across multiple programming languages and benchmarks demonstrate that EffiSkel significantly enhances both functional correctness and efficiency, resulting on Mercury with DeepSeek-Coder (7B) a +11.11% (vs. EffiCoder) and +3.71% (vs. CodeDPO) higher Efficiency Ratio (ER), and a +0.36 (vs. EffiCoder) and +0.22 (vs. CodeDPO) increase in Average Speedup (AS). These results highlight the effectiveness of explicitly modeling efficiency skeletons in improving the runtime performance of code generated by LLMs.

cs.SE↗

arXiv2026

Global adiabatic criterion for fast topological photon transfer in Fock-state lattices

Topological state transfer in Fock-state lattices has been demonstrated with high speed using sinusoidal profiles of coupling, yet the underlying reason has remained unclear. A global adiabatic criterion (GAC) is developed to bound the infidelity by the mean and variance of the nonadiabatic factor. The GAC reveals that the key to fast transfer is not a constant energy gap but the vanishing nonadiabaticity variance. For power-law coupling profiles, the variance vanishes only for the sinusoidal shape, which is thus globally optimal. Incorporating experimental decoherence parameters, it is predicted that the optimal transfer duration for a five-photon state is 161 ns, far shorter than 600 ns used in the experiment, reducing time by over 73% while increasing transferred photons by 29%. The optimal duration follow a simple linear scaling with photon number, providing a practical guideline. Through constructing an alternative constant-gap coupling family, it is confirmed that a constant gap alone is not sufficient for fast topological photon transfer. The essential condition is uniformity of nonadiabaticity. This work offers a rigorous explanation for the observed speed and a general framework for fast topological photonics engineering.

quant-ph↗

arXiv2026

Principled Synthetic Data Enables the First Scaling Laws for LLMs in Recommendation

Large Language Models (LLMs) represent a promising frontier for recommender systems, yet their development has been impeded by the absence of predictable scaling laws, which are crucial for guiding research and optimizing resource allocation. We hypothesize that this may be attributed to the inherent noise, bias, and incompleteness of raw user interaction data in prior continual pre-training (CPT) efforts. This paper introduces a novel, layered framework for generating high-quality synthetic data that circumvents such issues by creating a curated, pedagogical curriculum for the LLM. We provide powerful, direct evidence for the utility of our curriculum by showing that standard sequential models trained on our principled synthetic data significantly outperform ($+130\%$ on recall@100 for SasRec) models trained on real data in downstream ranking tasks, demonstrating its superiority for learning generalizable user preference patterns. Building on this, we empirically demonstrate, for the first time, robust power-law scaling for an LLM that is continually pre-trained on our high-quality, recommendation-specific data. Our experiments reveal consistent and predictable perplexity reduction across multiple synthetic data modalities. These findings establish a foundational methodology for reliable scaling LLM capabilities in the recommendation domain, thereby shifting the research focus from mitigating data deficiencies to leveraging high-quality, structured information.

cs.IR↗