arXiv subjects

Bei Yu

Publications and source records attributed to Bei Yu.

At least 55 records · Page 3Linked to original sources

arXiv2026

FocuSFT: Bilevel Optimization for Dilution-Aware Long-Context Fine-Tuning

Large language models can now process increasingly long inputs, yet their ability to effectively use information spread across long contexts remains limited. We trace this gap to how attention budget is spent during supervised fine-tuning (SFT) on long sequences: positional biases and attention sinks cause the model to allocate most of its attention to positionally privileged tokens rather than semantically relevant content. This training-time attention dilution (the starvation of content tokens in the attention distribution) weakens the gradient signal, limiting the model's ability to learn robust long-context capabilities. We introduce FocuSFT, a bilevel optimization framework that addresses this problem at training time. An inner loop adapts lightweight fast-weight parameters on the training context to form a parametric memory that concentrates attention on relevant content, and the outer loop performs SFT conditioned on this sharpened representation. Both loops apply bidirectional attention over context tokens while preserving causal masking for responses, reducing the causal asymmetry that gives rise to attention sinks and aligning inner-outer behavior. On BABILong, FocuSFT improves accuracy by up to +14pp across 4K--32K context lengths; on RULER, it raises CWE aggregation from 72.9\% to 81.1\% at 16K; and on GPQA with agentic tool use, it yields a 24\% relative gain in pass@1. Attention analysis shows that FocuSFT reduces attention sink mass by 529$\times$ and triples context engagement during training. Code: https://github.com/JarvisPei/FocuSFT

cs.CL↗

arXiv2026

VP-VLA: Visual Prompting as an Interface for Vision-Language-Action Models

Vision-Language-Action (VLA) models typically map visual observations and linguistic instructions directly to control signals. This "black-box" mapping forces a single forward pass to simultaneously handle instruction interpretation, spatial grounding, and low-level control, often leading to poor spatial precision and limited robustness in out-of-distribution scenarios. To address these limitations, we propose VP-VLA, a dual-system framework that decouples high-level reasoning and low-level execution via a structured visual prompting interface. Specifically, a "System 2 Planner" decomposes complex instructions into sub-tasks and identifies relevant target objects and goal locations. These spatial anchors are rendered directly within the native RGB observation space as modality-consistent visual prompts, such as crosshairs and bounding boxes. This avoids the modality mismatch introduced by dense masks, affordance maps, or additional control-specific representations. Guided by these prompts and enhanced by a novel auxiliary visual grounding objective during training, a "System 1 Controller" reliably generates precise low-level execution motions. Extensive experiments in simulation and real world demonstrate that VP-VLA surpasses state-of-the-art end-to-end baselines including QwenOFT and GR00T-N1.6. Project page: https://visualprompt-vla.github.io/

cs.RO↗

arXiv2026

UniMoCo: Unified Modality Completion for Robust Multi-Modal Embeddings

Current vision-language models have been explored for multi-modal embedding tasks like information retrieval. However, they face significant challenges in real-world queries and targets involving diverse modality combinations, as existing approaches often fail to align all modality combinations within a unified embedding space during training, leading to degraded performance on rare modality patterns during inference. To address this fundamental limitation, we propose UniMoCo, a novel architecture featuring a modality-completion module that generates visual features from text, thereby ensuring modality completeness for both queries and targets. Additionally, UniMoCo incorporates a specialized training strategy that aligns embeddings from both original and modality-completed inputs, thus ensuring consistent and robust embeddings for diverse modality combinations. Comprehensive experiments demonstrate that UniMoCo outperforms previous methods while exhibiting consistent robustness across diverse settings. Furthermore, we identify and quantify the inherent bias in conventional approaches caused by imbalanced modality combinations in training data, showing that our modality-completion paradigm effectively mitigates this limitation. The code is available at https://github.com/HobbitQia/UniMoCo.

cs.CV↗

arXiv2026

PipeRTL: Timing-Aware Pipeline Optimization at IR-Level for RTL Generation

Modern hardware compilers increasingly rely on rich intermediate representations (IRs) to preserve optimization-relevant semantics before generating RTL code. However, one important optimization is still largely deferred to backend tools: pipeline optimization. In common RTL flows, registers are inserted by frontend heuristics or hardware designers and later adjusted by backend retiming after the design has been lowered to a much lower-level netlist representation. At that point, much of the operator-level structure originally exposed by the compiler IR has already been weakened or lost, limiting opportunities for global, compiler-level pipeline optimization. This paper presents PipeRTL, an IR-level pipeline optimization framework for hardware compilers, instantiated in CIRCT. PipeRTL makes the legality of register relocation explicit in the IR, uses a learned timing predictor to approximate downstream delay behavior, and formulates timing-aware register relocation as a global min-cost flow problem under timing constraints. Evaluation on open-source designs under a commercial backend synthesis flow shows that PipeRTL improves downstream implementation quality on average, reducing critical-path delay, power, and area across the evaluated benchmarks, while also providing a stronger starting point for backend retiming. These results indicate that exposing pipeline optimization as an explicit compiler pass can deliver backend-meaningful gains by improving the sequential structure presented to later stages and the resulting downstream implementation quality.

cs.AR↗

arXiv2026

Graph World Models: Concepts, Taxonomy, and Future Directions

As one of the mainstream models of artificial intelligence, world models allow agents to learn the representation of the environment for efficient prediction and planning. However, classical world models based on flat tensors face several key problems, including noise sensitivity, error accumulation and weak reasoning. To address these limitations, many recent studies use graph structure to decompose the environment into entity nodes and interactive edges, and model virtual environments in a structured space. This paper systematically formalizes and unifies these emerging graph-based works under the concept of graph world models (GWMs). To the best of our knowledge, GWMs have not yet been explicitly defined and surveyed as a unified research paradigm. Furthermore, we propose a taxonomy based on relational inductive biases (RIB), categorizing GWMs by the specific structural priors they inject: (1) spatial RIB for topological abstraction; (2) physical RIB for dynamic simulation; and (3) logical RIB for causal and semantic reasoning. For each model category, we outline the key design principles, summarize representative models, and conduct comparative analyses. We further discuss open challenges and future directions, including dynamic graph adaptation, probabilistic relational dynamics, multi-granularity inductive biases, and the need for dedicated benchmarks and evaluation metrics for GWMs.

cs.AI↗

arXiv2026

Analytical FFN-to-MoE Restructuring via Activation Pattern Analysis

Scaling large language models (LLMs) improves performance but significantly increases inference costs, with feed-forward networks (FFNs) consuming the majority of computational resources. While Mixture-of-Experts (MoE) architectures can reduce this cost through sparse activation, restructuring existing dense models into MoEs typically requires extensive retraining on hundreds of billions of tokens. We propose an analytical post-training framework that rapidly restructures FFNs into sparse MoE architectures using only a small calibration dataset. The method analyzes neuron activation patterns to partition neurons into always-active shared experts and conditionally activated routed experts, then constructs a router analytically from representative neuron statistics, enabling immediate deployment or optional lightweight fine-tuning. This approach applies both to dense models and recursively to existing MoE models for hierarchical sparsity. Experiments demonstrate up to $1.17\times$ speedup in compute-bound scenarios with only minutes of processing and 2k-sample fine-tuning, outperforming methods requiring orders of magnitude more resources.

cs.LG↗

arXiv2026

MemDLM: Memory-Enhanced DLM Training

Diffusion Language Models (DLMs) offer attractive advantages over Auto-Regressive (AR) models, such as full-attention parallel decoding and flexible generation. However, standard DLM training uses a static, single-step masked prediction objective that never exposes the model to the progressive denoising dynamics of inference, and forces all contextual information to be maintained purely through token-space attention, which becomes increasingly diluted as context length grows. We propose MemDLM (Memory-Enhanced DLM), which introduces a second memory channel by embedding a simulated denoising trajectory into training via Bi-level Optimization. An inner loop updates a set of fast weights, forming a Parametric Memory that captures the local trajectory experience, while an outer loop updates the base model conditioned on this memory. By offloading part of the memorization burden from token-space attention to parameter space, MemDLM yields faster convergence, stronger long-context representations, and lower training loss, even when the fast weights are discarded at inference time. Re-enabling the inner loop at inference provides an additional prompt-specific adaptation effect, where the Parametric Memory acts as an emergent in-weight retrieval mechanism on challenging Needle-in-a-Haystack tasks. Code: https://github.com/JarvisPei/MemDLM.

cs.CL↗

arXiv2026

KCLNet: Electrically Equivalence-Oriented Graph Representation Learning for Analog Circuits

Digital circuits representation learning has made remarkable progress in the electronic design automation domain, effectively supporting critical tasks such as testability analysis and logic reasoning. However, representation learning for analog circuits remains challenging due to their continuous electrical characteristics compared to the discrete states of digital circuits. This paper presents a direct current (DC) electrically equivalent-oriented analog representation learning framework, named \textbf{KCLNet}. It comprises an asynchronous graph neural network structure with electrically-simulated message passing and a representation learning method inspired by Kirchhoff's Current Law (KCL). This method maintains the orderliness of the circuit embedding space by enforcing the equality of the sum of outgoing and incoming current embeddings at each depth, which significantly enhances the generalization ability of circuit embeddings. KCLNet offers a novel and effective solution for analog circuit representation learning with electrical constraints preserved. Experimental results demonstrate that our method achieves significant performance in a variety of downstream tasks, e.g., analog circuit classification, subcircuit detection, and circuit edit distance prediction.

cs.LG↗

arXiv2026

VisionZip: Longer is Better but Not Necessary in Vision Language Models

Recent advancements in vision-language models have enhanced performance by increasing the length of visual tokens, making them much longer than text tokens and significantly raising computational costs. However, we observe that the visual tokens generated by popular vision encoders, such as CLIP and SigLIP, contain significant redundancy. To address this, we introduce VisionZip, a simple yet effective method that selects a set of informative tokens for input to the language model, reducing visual token redundancy and improving efficiency while maintaining model performance. The proposed VisionZip can be widely applied to image and video understanding tasks and is well-suited for multi-turn dialogues in real-world scenarios, where previous methods tend to underperform. Experimental results show that VisionZip outperforms the previous state-of-the-art method by at least 5% performance gains across nearly all settings. Moreover, our method significantly enhances model inference speed, improving the prefilling time by 8x and enabling the LLaVA-Next 13B model to infer faster than the LLaVA-Next 7B model while achieving better results. Furthermore, we analyze the causes of this redundancy and encourage the community to focus on extracting better visual features rather than merely increasing token length. Our code is available at https://github.com/dvlab-research/VisionZip .

cs.CV↗

arXiv2026

Diversity or Precision? A Deep Dive into Next Token Prediction

Recent advancements have shown that reinforcement learning (RL) can substantially improve the reasoning abilities of large language models (LLMs). The effectiveness of such RL training, however, depends critically on the exploration space defined by the pre-trained model's token-output distribution. In this paper, we revisit the standard cross-entropy loss, interpreting it as a specific instance of policy gradient optimization applied within a single-step episode. To systematically study how the pre-trained distribution shapes the exploration potential for subsequent RL, we propose a generalized pre-training objective that adapts on-policy RL principles to supervised learning. By framing next-token prediction as a stochastic decision process, we introduce a reward-shaping strategy that explicitly balances diversity and precision. Our method employs a positive reward scaling factor to control probability concentration on ground-truth tokens and a rank-aware mechanism that treats high-ranking and low-ranking negative tokens asymmetrically. This allows us to reshape the pre-trained token-output distribution and investigate how to provide a more favorable exploration space for RL, ultimately enhancing end-to-end reasoning performance. Contrary to the intuition that higher distribution entropy facilitates effective exploration, we find that imposing a precision-oriented prior yields a superior exploration space for RL.

cs.CL↗

arXiv2026

SwingArena: Competitive Programming Arena for Long-context GitHub Issue Solving

We present SwingArena, a competitive evaluation framework for Large Language Models (LLMs) that closely mirrors real-world software development workflows. Unlike traditional static benchmarks, SwingArena models the collaborative process of software iteration by pairing LLMs as submitters, who generate patches, and reviewers, who create test cases and verify the patches through continuous integration (CI) pipelines. To support these interactive evaluations, we introduce a retrieval-augmented code generation (RACG) module that efficiently handles long-context challenges by providing syntactically and semantically relevant code snippets from large codebases, supporting multiple programming languages (C++, Python, Rust, and Go). This enables the framework to scale across diverse tasks and contexts while respecting token limitations. Our experiments, using over 400 high-quality real-world GitHub issues selected from a pool of 2,300 issues, show that models like GPT-4o excel at aggressive patch generation, whereas DeepSeek and Gemini prioritize correctness in CI validation. SwingArena presents a scalable and extensible methodology for evaluating LLMs in realistic, CI-driven software development settings. More details are available on our project page: swing-bench.github.io

cs.CL↗

arXiv2026

Low-Light Video Enhancement with An Effective Spatial-Temporal Decomposition Paradigm

Low-Light Video Enhancement (LLVE) seeks to restore dynamic or static scenes plagued by severe invisibility and noise. In this paper, we present an innovative video decomposition strategy that incorporates view-independent and view-dependent components to enhance the performance of LLVE. The framework is called View-aware Low-light Video Enhancement (VLLVE). We leverage dynamic cross-frame correspondences for the view-independent term (which primarily captures intrinsic appearance) and impose a scene-level continuity constraint on the view-dependent term (which mainly describes the shading condition) to achieve consistent and satisfactory decomposition results. To further ensure consistent decomposition, we introduce a dual-structure enhancement network featuring a cross-frame interaction mechanism. By supervising different frames simultaneously, this network encourages them to exhibit matching decomposition features. This mechanism can seamlessly integrate with encoder-decoder single-frame networks, incurring minimal additional parameter costs. Building upon VLLVE, we propose a more comprehensive decomposition strategy by introducing an additive residual term, resulting in VLLVE++. This residual term can simulate scene-adaptive degradations, which are difficult to model using a decomposition formulation for common scenes, thereby further enhancing the ability to capture the overall content of videos. In addition, VLLVE++ enables bidirectional learning for both enhancement and degradation-aware correspondence refinement (end-to-end manner), effectively increasing reliable correspondences while filtering out incorrect ones. Notably, VLLVE++ demonstrates strong capability in handling challenging cases, such as real-world scenes and videos with high dynamics. Extensive experiments are conducted on widely recognized LLVE benchmarks.

cs.CV↗

arXiv2026

VisionReasoner: Unified Reasoning-Integrated Visual Perception via Reinforcement Learning

Large vision-language models exhibit inherent capabilities to handle diverse visual perception tasks. In this paper, we introduce VisionReasoner, a unified framework capable of reasoning and solving multiple visual perception tasks within a shared model. Specifically, by designing a unified reward mechanism and multi-object cognitive learning strategies, VisionReasoner enhances its reasoning capabilities to analyze visual inputs, and addresses diverse perception tasks within a unified model. VisionReasoner generates a structured reasoning process before delivering the desired outputs responding to user queries. Human evaluation reveals the reasoning process of VisionReasoner is faithful and reliable even without annotated reasoning train data. To rigorously assess unified visual perception capabilities, we evaluate VisionReasoner on ten diverse tasks spanning three critical domains: detection, segmentation, and counting. Experimental results show that VisionReasoner achieves superior performance as a unified model, outperforming the baseline Qwen2.5VL by relative margins of 29.1\% on COCO (detection), 22.1\% on ReasonSeg (segmentation), and 13.2\% on CountBench (counting).

cs.CV↗

arXiv2026

One-Token Rollout: Guiding Supervised Fine-Tuning of LLMs with Policy Gradient

Supervised fine-tuning (SFT) is the predominant method for adapting large language models (LLMs), yet it often struggles with generalization compared to reinforcement learning (RL). In this work, we posit that this performance disparity stems not just from the loss function, but from a more fundamental difference: SFT learns from a fixed, pre-collected dataset, whereas RL utilizes on-policy data sampled from the current policy. Building on this hypothesis, we introduce one-token rollout (OTR), a novel fine-tuning algorithm that guides SFT with the policy gradient method. OTR reframes the autoregressive learning process by treating each token generation as a single-step reinforcement learning trajectory. At each step, it performs a Monte Carlo ``rollout'' by sampling multiple candidate tokens from the current policy's distribution. The ground-truth token from the supervised data is then used to provide a reward signal to these samples. Guided by policy gradient, our algorithm repurposes static, off-policy supervised data into a dynamic, on-policy signal at the token level, capturing the generalization benefits of on-policy learning while bypassing the costly overhead of full sentence generation. Through extensive experiments on a diverse suite of challenging benchmarks spanning mathematical reasoning, code generation, and general domain reasoning, we demonstrate that OTR consistently outperforms standard SFT. Our findings establish OTR as a powerful and practical alternative for fine-tuning LLMs and provide compelling evidence that the on-policy nature of data is a critical driver of generalization, offering a promising new direction for fine-tuning LLMs.

cs.CL↗

arXiv2025

ToTRL: Unlock LLM Tree-of-Thoughts Reasoning Potential through Puzzles Solving

Large language models (LLMs) demonstrate significant reasoning capabilities, particularly through long chain-of-thought (CoT) processes, which can be elicited by reinforcement learning (RL). However, prolonged CoT reasoning presents limitations, primarily verbose outputs due to excessive introspection. The reasoning process in these LLMs often appears to follow a trial-and-error methodology rather than a systematic, logical deduction. In contrast, tree-of-thoughts (ToT) offers a conceptually more advanced approach by modeling reasoning as an exploration within a tree structure. This reasoning structure facilitates the parallel generation and evaluation of multiple reasoning branches, allowing for the active identification, assessment, and pruning of unproductive paths. This process can potentially lead to improved performance and reduced token costs. Building upon the long CoT capability of LLMs, we introduce tree-of-thoughts RL (ToTRL), a novel on-policy RL framework with a rule-based reward. ToTRL is designed to guide LLMs in developing the parallel ToT strategy based on the sequential CoT strategy. Furthermore, we employ LLMs as players in a puzzle game during the ToTRL training process. Solving puzzle games inherently necessitates exploring interdependent choices and managing multiple constraints, which requires the construction and exploration of a thought tree, providing challenging tasks for cultivating the ToT reasoning capability. Our empirical evaluations demonstrate that our ToTQwen3-8B model, trained with our ToTRL, achieves significant improvement in performance and reasoning efficiency on complex reasoning tasks.

cs.CL↗

arXiv2025

From Pruning to Grafting: Dynamic Knowledge Redistribution via Learnable Layer Fusion

Structured pruning of Generative Pre-trained Transformers (GPTs) offers a promising path to efficiency but often suffers from irreversible performance degradation due to the discarding of transformer blocks. In this paper, we introduce FuseGPT, a compression paradigm that reframes structured pruning as iterative knowledge grafting rather than simple removal. Motivated by the observation that linear block merging fails to capture non-linear feature disparities and that block importance fluctuates dynamically during pruning, FuseGPT employs a dual-strategy pipeline. First, we propose Macro Influence (MI), a dynamic fusion-aware metric that continuously re-evaluates block redundancy as the network topology evolves. Second, instead of rigid parameter averaging, we introduce a learnable low-rank fusion mechanism that adaptively grafts the knowledge of pruned blocks onto surviving layers via lightweight local distillation. Extensive experiments on LLaMA, Mistral, Qwen, and Phi families demonstrate that FuseGPT establishes a new state-of-the-art on the compression-accuracy Pareto frontier: at 25\% sparsity, FuseGPT achieves lower perplexity than prior methods at 20\% sparsity, improves zero-shot reasoning by up to 4.5 points, and delivers 1.33$\times$ inference speedup with 25\% memory reduction. Furthermore, FuseGPT is orthogonal to quantization, achieving 52.1\% total compression with negligible quality loss when combined with 4-bit GPTQ. We make our code publicly available at https://github.com/JarvisPei/FuseGPT.

cs.LG↗

arXiv2025

Bounded Dynamic Level Maintenance for Efficient Logic Optimization

Logic optimization constitutes a critical phase within the Electronic Design Automation (EDA) flow, essential for achieving desired circuit power, performance, and area (PPA) targets. These logic circuits are typically represented as Directed Acyclic Graphs (DAGs), where the structural depth, quantified by node level, critically correlates with timing performance. Modern optimization strategies frequently employ iterative, local transformation heuristics (\emph{e.g.,} \emph{rewrite}, \emph{refactor}) directly on this DAG structure. As optimization continuously modifies the graph locally, node levels require frequent dynamic updates to guide subsequent decisions. However, a significant gap exists: existing algorithms for incrementally updating node levels are unbounded to small changes. This leads to a total of worst complexity in $O(|V|^2)$ for given local subgraphs $\{ΔG_i\}_{i=1}^{|V|}$ updates on DAG $G(V,E)$. This unbounded nature poses a severe efficiency bottleneck, hindering the scalability of optimization flows, particularly when applied to large circuit designs prevalent today. In this paper, we analyze the dynamic level maintenance problem endemic to iterative logic optimization, framing it through the lens of partial topological order. Building upon the analysis, we present the first bounded algorithm for maintaining level constraints, with $O(|V| Δ\log Δ)$ time for a sequence $|V|$ of updates $\{ΔG_i\}$, where $Δ= \max_i \|ΔG_i\|$ denotes the maximum extended size of $ΔG_i$. Experiments on comprehensive benchmarks show our algorithm enables an average 6.4$\times$ overall speedup relative to \rw and \rf, driven by a 1074.8$\times$ speedup in the level maintenance, all without any quality sacrifice.

cs.CC↗

arXiv2025

MOSS: Efficient and Accurate FP8 LLM Training with Microscaling and Automatic Scaling

Training large language models with FP8 formats offers significant efficiency gains. However, the reduced numerical precision of FP8 poses challenges for stable and accurate training. Current frameworks preserve training performance using mixed-granularity quantization, i.e., applying per-group quantization for activations and per-tensor/block quantization for weights. While effective, per-group quantization requires scaling along the inner dimension of matrix multiplication, introducing additional dequantization overhead. Moreover, these frameworks often rely on just-in-time scaling to dynamically adjust scaling factors based on the current data distribution. However, this online quantization is inefficient for FP8 training, as it involves multiple memory reads and writes that negate the performance benefits of FP8. To overcome these limitations, we propose MOSS, a novel FP8 training framework that ensures both efficiency and numerical stability. MOSS introduces two key innovations: (1) a two-level microscaling strategy for quantizing sensitive activations, which balances precision and dequantization cost by combining a high-precision global scale with compact, power-of-two local scales; and (2) automatic scaling for weights in linear layers, which eliminates the need for costly max-reduction operations by predicting and adjusting scaling factors during training. Leveraging these techniques, MOSS enables efficient FP8 training of a 7B parameter model, achieving performance comparable to the BF16 baseline while achieving up to 34% higher training throughput.

cs.LG↗