Searcharxiv⌕ Search

arXiv subjects

Siyuan Lu

Publications and source records attributed to Siyuan Lu.

At least 19 recordsLinked to original sources

No Second Run: Does RSI Stop at Frontier Pre-training?

Pre-training is where every foundation model begins; bringing recursive self-improvement (RSI) there would let AI improve the process that creates AI. Yet public RSI work has so far targeted only post-training, which starts from an existing base model, runs trials costing GPU-hours, and returns verifiable scores. Pre-training breaks all three conditions at orders-of-magnitude higher cost: a run consumes hundreds of thousands of GPU-hours over months and cannot be repeated; its objective, post-training headroom, cannot be read off any single score; and what must improve is not a model but the whole production system, from data pipelines and training infrastructure to the teams that run them. To our knowledge, we are the first to apply RSI to an entire frontier pre-training run from scratch and to report it in detail. RSI does not stop at frontier pre-training, but it changes form: since no end-to-end loop can close over a single unrepeatable run, we decompose it into smaller loops that can, each repairing a step the full run breaks. The result is 801-30B-A3B-Base, a 30B-A3B MoE trained from scratch on 11.8T tokens in three months by a newly formed team. We release it as a base model; the lightweight SFT reported here (6B tokens, no RL) only probes post-training headroom. As a base model, 801-30B-A3B outperforms Qwen3-30B-A3B and Nemotron-3.5-Lightning-30B-A3B, trailing only the larger Qwen3.5-35B-A3B, and keeps this position under identical SFT; with lightweight SFT alone, it achieves the highest average on four agentic benchmarks against fully post-trained models, including Qwen3.6-35B-A3B. Since post-training RSI presupposes a strong base model, bringing RSI into pre-training acts on the foundation every later stage builds on. We open-source 801-30B-A3B-Base and the lightweight SFT data at https://github.com/GAIR-NLP/801-30B .

cs.AI↗

CacheBridge: Efficient Cross-Model KV Transfer for Multi-Model LLM Serving

Multi-model LLM serving uses routers, cascades and agent workflows to move requests across models with different costs and capabilities. Existing serving systems reuse KV caches only within one model, so every switch forces the target to re-prefill history the source already processed, adding time to first token (TTFT) that grows with history length and target size. Cross-model KV transfer can replace this prefill, yet prior methods are slower than re-prefill end to end. We present CacheBridge, the first serving system to accelerate heterogeneous multi-model workloads through cross-model KV transfer. It co-designs a compact, attention-calibrated KV mapper with systematic optimization for online multi-model serving, which makes model switch efficient on latency and memory cost. In Qwen3 14B$\rightarrow$32B agent serving on $τ$-bench, it reduces mean TTFT by $30$ to $98\%$ versus re-prefill at $1$ to $24$ QPS with comparable task success. Across five model pairs from three families, it uses $8$ to $12\times$ less mapper memory than Closed-Form Mapper and retains on average $97.94\%$ and $102.94\%$ of re-prefill task quality on chat/reasoning and RAG workloads, respectively. The promising results have opened new opportunities for efficient multi-model LLM serving through cross-model KV reuse.

cs.AI↗

Trajectory2Gym: Decision-Grounded Compositional Interaction Synthesis for Tool-Use Agents

Successful tool-use trajectories are not necessarily useful supervision. Even successful interactions may underrepresent multi-step dependencies or contain actions that are unjustified by the information visible to the agent at decision time. We identify two coupled requirements for effective trajectory supervision: compositional coverage, so that important observation-action dependencies are represented, and decision grounding, so that each supervised action is supported by the learner-visible interaction history. We introduce Trajectory2Gym, which constructs executable interactions around under-supported compositional structures and derives supervision only from grounded decisions in their realized executions. Structural-support selection allocates the resulting experience, while parent-normalized supervised fine-tuning preserves the semantic interaction as the unit of training weight. This reframes trajectory generation as the deliberate construction of learnable interaction experience, rather than the collection of successful trajectories. Across five Qwen3 scales, Trajectory2Gym improves Tau1, Tau2, and BFCL at every tested scale, increasing the equal-weight three-benchmark mean by 16.9-29.8 percentage points. The gains are strongest in stateful and multi-turn interaction, where later decisions depend on earlier observations and feedback. Our results suggest that scaling tool-use agents depends not only on collecting more successful trajectories, but on constructing interactions whose dependencies are represented and whose decisions are learnable from the agent's visible history.

cs.CL↗

Teach the Magnitude, Not the Direction: Verifier-Anchored Credit Assignment for Multi-Turn Multi-Step LLM Agents

Reinforcement learning with verifiable rewards (RLVR) directly optimizes verifier-defined outcomes for multi-turn tool-use agents, yet its trajectory-level credit assignment conflates heterogeneous per-turn outcomes into a single reward signal. On-policy distillation provides dense per-token supervision but uses a teacher-defined token-level objective and can be prone to gradient concentration collapse. We introduce $\textbf{CrEST}$, a hierarchical credit assignment framework that anchors each sampled token's scalar advantage sign to verified rewards while incorporating dense magnitude signals from a privileged self-teacher. $\textbf{CrEST}$ resolves credit at two levels: turn-segmented verified advantages address inter-turn dilution, while entropy-gated self-teacher modulation refines intra-turn token contributions. Experiments on BFCL V3 and WildToolBench show that $\textbf{CrEST}$ consistently outperforms both RL and distillation baselines across two model scales, with the largest gains on long-trajectory and strict session-level metrics. Our work demonstrates that the teacher's role in policy optimization can be reduced from determining token-level reinforcement or suppression to modulating update magnitudes, enabling dense credit assignment while keeping verifier rewards as the training objective.

cs.AI↗

RODS: Reward-Driven Online Data Synthesis for Multi-Turn Tool-Use Agents

Multi-turn tool-use RL is bottlenecked by the rapid depletion of informative samples in static datasets. We observe that the gradient signal in GRPO concentrates on tasks with the highest rollout reward variance, a consequence of the Popoviciu upper bound. Consequently, samples near the agent's capability boundary -- where successes and failures are roughly balanced -- contribute disproportionately large policy gradients. As training progresses, this boundary continuously shifts, which gradually depletes the pool of informative samples in a static dataset. We propose RODS (Reward-driven Online Data Synthesis) to resolve this depletion. RODS closes the loop between RL training and data generation by repurposing the progress reward variance as a practical, zero-cost boundary detector that requires no extra inference beyond the rollouts already computed for training. It continuously identifies such boundary samples, synthesizes new multi-turn variants matching their structural complexity (e.g., API topology and dependency depth) via a skill-aligned resampling pipeline, and manages a dynamic replay buffer that co-evolves with the policy. Starting from 400 human seeds and maintaining an active training pool of ~800 samples, RODS achieves comparable performance to a 17K-sample offline pipeline while requiring roughly 20x fewer trajectories, and improves over fixed-data RL and environment augmentation in our controlled setting.

cs.AI↗

X-Cache: Cross-Chunk Block Caching for Few-Step Autoregressive World Models Inference

Real-time world simulation is becoming a key infrastructure for scalable evaluation and online reinforcement learning of autonomous driving systems. Recent driving world models built on autoregressive video diffusion achieve high-fidelity, controllable multi-camera generation, but their inference cost remains a bottleneck for interactive deployment. However, existing diffusion caching methods are designed for offline video generation with multiple denoising steps, and do not transfer to this scenario. Few-step distilled models have no inter-step redundancy left for these methods to reuse, and sequence-level parallelization techniques require future conditioning that closed-loop interactive generation does not provide. We present X-Cache, a training-free acceleration method that caches along a different axis: across consecutive generation chunks rather than across denoising steps. X-Cache maintains per-block residual caches that persist across chunks, and applies a dual-metric gating mechanism over a structure- and action-aware block-input fingerprint to independently decide whether each block should recompute or reuse its cached residual. To prevent approximation errors from permanently contaminating the autoregressive KV cache, X-Cache identifies KV update chunks (the forward passes that write clean keys and values into the persistent cache) and unconditionally forces full computation on these chunks, cutting off error propagation. We implement X-Cache on X-world, a production multi-camera action-conditioned driving world model built on multi-block causal DiT with few-step denoising and rolling KV cache. X-Cache achieves 71% block skip rate with 2.6x wall-clock speedup while maintaining minimum degradation.

cs.CV↗

Federated Reasoning Distillation Framework with Model Learnability-Aware Data Allocation

Data allocation plays a critical role in federated large language model (LLM) and small language models (SLMs) reasoning collaboration. Nevertheless, existing data allocation methods fail to address an under-explored challenge in collaboration: bidirectional model learnability gap, where client-side SLMs cannot identify high-reward samples matching their learnability constraints for effective knowledge transfer from LLMs, while LLMs struggle to select samples contributing novel knowledge beyond their existing data. Furthermore, these collaboration frameworks face another key challenge: domain-agnostic reasoning transfer, where existing reasoning transfer methods fail to flexibly adapt to the local domain data, preventing SLMs from effectively acquiring step-by-step reasoning abilities within from general LLM. To address these challenges, we propose LaDa, a federated reasoning distillation framework with model learnability-aware data allocation. It introduces a model learnability-aware data filter that adaptively allocates high-reward samples based on the learnability gap between each SLM and LLM pair, effectively facilitating bidirectional knowledge transfer. We further design a domain adaptive reasoning distillation method that aligns joint probabilities of reasoning paths on filtered high-reward samples through contrastive distillation learning between SLM and LLM, enabling SLM to capture underlying reasoning patterns under local data distribution. LaDa operates as a plug-in module for existing collaboration frameworks, adapting knowledge transfer based on model learnability gaps.

cs.AI↗

Dual-Kernel Adapter: Expanding Spatial Horizons for Data-Constrained Medical Image Analysis

Adapters have become a widely adopted strategy for efficient fine-tuning of large pretrained models, particularly in resource-constrained settings. However, their performance under extreme data scarcity, common in medical imaging due to high annotation costs, privacy regulations, and fragmented datasets, remains underexplored. In this work, we present the first comprehensive study of adapter-based fine-tuning for large pretrained models in low-data medical imaging scenarios. We find that, contrary to their promise, conventional adapters can degrade performance under severe data constraints, performing even worse than simple linear probing when trained on less than 1% of the corresponding training data. Through systematic analysis, we identify a sharp reduction in Effective Receptive Field (ERF) as a key factor behind this degradation. Motivated by these findings, we propose the Dual-Kernel Adapter (DKA), a lightweight module that expands spatial context via large-kernel convolutions while preserving local detail with small-kernel counterparts. Extensive experiments across diverse classification and segmentation benchmarks show that DKA significantly outperforms existing adapter methods, establishing new leading results in both data-constrained and data-rich regimes.

cs.CE↗

On Liouville's theorem for the Hessian quotient equation $σ_2/σ_1$

We prove Liouville's theorem for semi-convex entire solutions to Hessian quotient equation $σ_2/σ_1=1$ in $\mathbb{R}^n$. The proof is based on the observation that after rewriting the quotient operator as the $σ_2$ operator, acting on a new function, one can refer to the recent result of Shankar and Yuan on Liouville's theorem for $σ_2$ equation.

math.AP↗

Don't Just Fine-tune the Agent, Tune the Environment

Large Language Model (LLM) agents show great promise for complex, multi-turn tool-use tasks, but their development is often hampered by the extreme scarcity of high-quality training data. Supervised fine-tuning (SFT) on synthetic data leads to overfitting, whereas standard reinforcement learning (RL) struggles with a critical cold-start problem and training instability. To address these challenges, we introduce $\textbf{Environment Tuning}$, a novel training paradigm that enables agents to learn complex behaviors directly from problem instances without relying on pre-collected expert trajectories. $\textbf{Environment Tuning}$ orchestrates this learning process through a structured curriculum, actionable environment augmentation that provides corrective feedback, and fine-grained progress rewards to ensure stable and efficient exploration. Using only 400 problem instances from Berkeley Function-Calling Leaderboard (BFCL) benchmark, our method not only achieves competitive in-distribution performance against strong baselines but also demonstrates superior out-of-distribution generalization, overcoming the performance collapse common to SFT-based approaches. Our work presents a paradigm shift from supervised fine-tuning on static trajectories to dynamic, environment-based exploration, paving the way for training more robust and data-efficient agents. The code is available at https://github.com/inclusionAI/AWorld-RL/tree/main/EnvTuning.

cs.AI↗

HADUA: Hierarchical Attention and Dynamic Uniform Alignment for Robust Cross-Subject Emotion Recognition

Robust cross-subject emotion recognition from multimodal physiological signals remains a challenging problem, primarily due to modality heterogeneity and inter-subject distribution shift. To tackle these challenges, we propose a novel adaptive learning framework named Hierarchical Attention and Dynamic Uniform Alignment (HADUA). Our approach unifies the learning of multimodal representations with domain adaptation. First, we design a hierarchical attention module that explicitly models intra-modal temporal dynamics and inter-modal semantic interactions (e.g., between electroencephalogram(EEG) and eye movement(EM)), yielding discriminative and semantically coherent fused features. Second, to overcome the noise inherent in pseudo-labels during adaptation, we introduce a confidence-aware Gaussian weighting scheme that smooths the supervision from target-domain samples by down-weighting uncertain instances. Third, a uniform alignment loss is employed to regularize the distribution of pseudo-labels across classes, thereby mitigating imbalance and stabilizing conditional distribution matching. Extensive experiments on multiple cross-subject emotion recognition benchmarks show that HADUA consistently surpasses existing state-of-the-art methods in both accuracy and robustness, validating its effectiveness in handling modality gaps, noisy pseudo-labels, and class imbalance. Taken together, these contributions offer a practical and generalizable solution for building robust cross-subject affective computing systems.

cs.MM↗

SDC-Net: A Domain Adaptation Framework with Semantic-Dynamic Consistency for Cross-Subject EEG Emotion Recognition

Emotion recognition based on electroencephalography (EEG) holds significant promise for affective brain-computer interfaces (aBCIs). However, its practical deployment faces challenges due to the variability within inter-subject and the scarcity of labeled data in target domains. To overcome these limitations, we propose SDC-Net, a novel Semantic-Dynamic Consistency domain adaptation network for fully label-free cross-subject EEG emotion recognition. First, we introduce a Same-Subject Same-Trial Mixup strategy that generates augmented samples through intra-trial interpolation, enhancing data diversity while explicitly preserving individual identity to mitigate label ambiguity. Second, we construct a dynamic distribution alignment module within the Reproducing Kernel Hilbert Space (RKHS), jointly aligning marginal and conditional distributions through multi-objective kernel mean embedding, and leveraging a confidence-aware pseudo-labeling strategy to ensure stable adaptation. Third, we propose a dual-domain similarity consistency learning mechanism that enforces cross-domain structural constraints based on latent pairwise similarities, facilitating semantic boundary learning without reliance on temporal synchronization or label priors. To validate the effectiveness and robustness of the proposed SDC-Net, extensive experiments are conducted on three widely used EEG benchmark datasets: SEED, SEED-IV, and FACED. Comparative results against existing unsupervised domain adaptation methods demonstrate that SDC-Net achieves state-of-the-art performance in emotion recognition under both cross-subject and cross-session conditions. This advancement significantly improves the accuracy and generalization capability of emotion decoding, laying a solid foundation for real-world applications of personalized aBCIs. The source code is available at: https://github.com/XuanSuTrum/SDC-Net.

cs.HC↗

Testing and Fault Tolerance Techniques for CNT-Based FPGAs

As the semiconductor manufacturing process technology node shrinks into the nanometer-scale, the CMOS-based Field Programmable Gate Arrays (FPGAs) face big challenges in scalability of performance and power consumption. Multi-walled Carbon Nanotube (MWCNT) serves as a promising candidate for Cu interconnects thanks to the superior conductivity. Moreover, Carbon Nanotube Field Transistor (CNFET) also emerges as a prospective alternative to the conventional CMOS device because of high power efficiency and large noise margin. The combination of MWCNT and CNFET enables the promising CNT-based FPGAs. However, the MWCNT interconnects exhibit significant process variations due to immature fabrication process, leading to delay faults. Also, the non-ideal CNFET fabrication process may generate a few metallic CNTs (m-CNTs), rendering correlated faulty blocks. In this article, we propose a ring oscillator (RO) based testing technique to detect delay faults due to the process variation of MWCNT interconnects. Furthermore, we propose an effective testing technique for the carry chains in CLBs, and an improved circuit design based on the lookup table (LUT) is applied to speed up the fault testing of CNT-based FPGAs. In addition, we propose a testing algorithm to detect m-CNTs in CLBs. Finally, we propose a redundant spare row sharing architecture to improve the yield of CNT-based FPGA further. Experimental results show that the test time for a 6-input LUT can be reduced by 35.49% compared with conventional testing, and the proposed algorithm can achieve a high test coverage with little overhead. The proposed redundant architecture can repair the faulty segment effectively and efficiently.

cs.AR↗

MorphAgent: Empowering Agents through Self-Evolving Profiles and Decentralized Collaboration

Large Language Model (LLM) based multi-agent systems (MAS) have shown promise in tackling complex tasks, but often rely on predefined roles and centralized coordination, limiting their adaptability to evolving challenges. This paper introduces MorphAgent, a novel Autonomous, Self-Organizing, and Self-Adaptive Multi-Agent System for decentralized agent collaboration that enables agents to dynamically evolve their roles and capabilities. Our approach employs self-evolving agent profiles, optimized through three key metrics, guiding agents in refining their individual expertise while maintaining complementary team dynamics. MorphAgent implements a two-phase process: a Profile Update phase for profile optimization, followed by a Task Execution phase where agents continuously adapt their roles based on task feedback. Our experimental results show that MorphAgent outperforms existing frameworks in terms of task performance and adaptability to changing requirements, paving the way for more robust and versatile multi-agent collaborative systems.

cs.AI↗

AWorld: Orchestrating the Training Recipe for Agentic AI

The learning from practice paradigm is crucial for developing capable Agentic AI systems, yet it is severely hampered by inefficient experience generation, a bottleneck especially pronounced in complex benchmarks like GAIA. To address this, we introduce AWorld, an open-source system engineered for large-scale agent-environment interaction. By distributing tasks across a cluster, AWorld accelerates experience collection by 14.6x compared to standard single-node, sequential execution. This critical speedup makes extensive reinforcement learning practical and scalable. Leveraging this capability, we trained a Qwen3-32B-based agent that achieves pass@1 accuracy of 32.23% on the GAIA test set, which surpasses GPT-4o (27.91%) and rivals DeepSeek-V3 (31.89%). Our open-source system and the resulting agent provide a practical blueprint for a complete agentic AI training pipeline, from efficient interaction to demonstrable model improvement.

cs.AI↗

You Don't Know Until You Click:Automated GUI Testing for Production-Ready Software Evaluation

Large Language Models (LLMs) and code agents in software development are rapidly evolving from generating isolated code snippets to producing full-fledged software applications with graphical interfaces, interactive logic, and dynamic behaviors. However, current benchmarks fall short in evaluating such production-ready software, as they often rely on static checks or binary pass/fail scripts, failing to capture the interactive behaviors and runtime dynamics that define real-world usability - qualities that only emerge when an application is actively used. This is the blind spot of current evaluation: you don't know if an app works until you click through it, interact with it, and observe how it responds. To bridge this gap, we introduce RealDevWorld, a novel evaluation framework for automated end-to-end assessment of LLMs' ability to generate production-ready repositories from scratch. It features two key components: (1) RealDevBench, a diverse collection of 194 open-ended software engineering tasks across multiple domains, incorporating multimodal elements to reflect real-world complexity; and (2) AppEvalPilot, a new agent-as-a-judge evaluation system that simulates realistic, GUI-based user interactions to automatically and holistically assess software functional correctness, visual fidelity, and runtime behavior. The framework delivers fine-grained, task-specific diagnostic feedback, supporting nuanced evaluation beyond simple success/failure judgments. Empirical results show that RealDevWorld delivers effective, automatic, and human-aligned evaluations, achieving an accuracy of 0.92 and a correlation of 0.85 with expert human assessments, while significantly reducing the reliance on manual review. This enables scalable, human-aligned assessment of production-level software generated by LLMs. Our code is available on GitHub.

cs.SE↗

H2Tune: Federated Foundation Model Fine-Tuning with Hybrid Heterogeneity

Different from existing federated fine-tuning (FFT) methods for foundation models, hybrid heterogeneous federated fine-tuning (HHFFT) is an under-explored scenario where clients exhibit double heterogeneity in model architectures and downstream tasks. This hybrid heterogeneity introduces two significant challenges: 1) heterogeneous matrix aggregation, where clients adopt different large-scale foundation models based on their task requirements and resource limitations, leading to dimensional mismatches during LoRA parameter aggregation; and 2) multi-task knowledge interference, where local shared parameters, trained with both task-shared and task-specific knowledge, cannot ensure only task-shared knowledge is transferred between clients. To address these challenges, we propose H2Tune, a federated foundation model fine-tuning with hybrid heterogeneity. Our framework H2Tune consists of three key components: (i) sparsified triple matrix decomposition to align hidden dimensions across clients through constructing rank-consistent middle matrices, with adaptive sparsification based on client resources; (ii) relation-guided matrix layer alignment to handle heterogeneous layer structures and representation capabilities; and (iii) alternating task-knowledge disentanglement mechanism to decouple shared and specific knowledge of local model parameters through alternating optimization. Theoretical analysis proves a convergence rate of O(1/\sqrt{T}). Extensive experiments show our method achieves up to 15.4% accuracy improvement compared to state-of-the-art baselines. Our code is available at https://anonymous.4open.science/r/H2Tune-1407.

cs.LG↗

Superradiance Enhanced Light-Matter Interaction in Spatially Ordered Shape and Volume Controlled Single Quantum Dots: Enabling On-Chip Photonic Networks

On-chip photonic networks require adequately spatially ordered matter-photon interconversion qubit sources with emission figures-of-merit exceeding the requirements that would enable the desired functional response of the network. The mesa-top single quantum dots (MTSQDs) have recently been demonstrated to meet these requirements. The substrate-encoded size-reducing epitaxy (SESRE) approach underpinning the realization of these unique quantum emitters allows control on the shape, size, and strain (lattice-matched or mismatched) of these epitaxial single quantum dots. We have exploited this unique feature of the MTSQDs to reproducibly create arrays of quantum dots that exhibit single photon superradiance, a characteristic of the SESRE-enabled delicate balance between the confinement pontential volume, depth, the resulting exciton binding energy, and the degree of confinement of the center of mass (CM) motion of the exciton. Scanning transmission electron microscope (STEM) studies reveal the structural (atomic scale) and chemical (nm scale) nature of the material region defining the notion of the shape and volume (here large) of the electron confinement region (i.e. the QD). In the exciton's weake CM confinement regime, owing to its coherent sampling of the large volume, enhancement of the MTSQD oscillator strength to ~30 is demonstrated. Theoretical modelling with input from the STEM finding provide corroboration for single photon superradiance causing enhancement of the oscillator strength by ~2.5 to 3. Our findings allow fabricating and studying interconnected networks enabled by the unique matter qubit-light qubit interconversion units that can be realized for lattice matched and mismatched material combinations covering UV to midinfrared wavelength range.

cond-mat.mes-hall↗