arXiv subjects

Han Shi

Publications and source records attributed to Han Shi.

At least 19 recordsLinked to original sources

arXiv2026

Budgeted KV-Cache Repair for Evolving Contexts

Language-model agents revise their working context as information becomes obsolete, but deleting text does not erase its influence on cached computation. Later key-value (KV) states remain conditioned on the deleted content, while recomputing the entire retained suffix can be costly. We introduce budgeted KV-cache repair: recovering task performance by recomputing a user-chosen fraction of the retained suffix. Comparing repair allocations under identical relevance scores and token budgets, we find that refreshing coherent regions can substantially outperform refreshing isolated tokens. This finding motivates Semantic-Unit Refresh (SUR), which allocates the budget to query-relevant text regions using inexpensive embedding scores, without training or an auxiliary transformer forward. To execute these repairs efficiently, a layer-packed executor processes selected states together at each layer while preserving full causal cache access. Across five models, SUR achieves strong recovery in controlled deletion tasks while recomputing only 2-5% of the suffix. On four natural multi-hop QA datasets, it achieves near-full-recompute exact match at a 5% budget with lower post-edit latency. The executor accelerates fragmented refresh by up to 16.1x over interval-wise execution.

cs.AI↗

arXiv2026

AvoKV-E: Payload-Aware KV Cache Eviction for Long Reasoning

Long-output reasoning shifts the KV-cache bottleneck from the fixed prompt to the generated trace. Existing reasoning-cache eviction methods largely treat cached entries as routing objects, estimating whether an old key will still be read, will recur, or can be replaced. This routing-only view overlooks two effects: low-attention entries can carry large value payloads whose removal changes future predictions, and newly generated states can appear stale before later queries have had a chance to read them. We introduce AvoKV-E, a training-free eviction policy that first delays eligibility for recent states and then ranks eligible entries using candidate-normalized read pressure, key redundancy, and value-payload potential. According to empirical evaluation across different models and datasets, AvoKV-E matches or exceeds redundancy-aware, recurrence-based, and thought-adaptive eviction baselines at matched active-KV budgets, with its largest gains in the tightest-cache regime. Component and counterfactual analyses further connect these gains to delayed observation, payload-aware scoring, redundancy, and scale-robust normalization. Together, the results show that long-reasoning KV eviction should preserve not only keys that are likely to be read, but also the value payloads that sustain the reasoning trajectory.

cs.LG↗

arXiv2026

UniPool: Learning Expert-to-Layer Ownership from Brief Global Access

Most Mixture-of-Experts (MoE) transformers preassign each expert to one layer for the whole of training. We study whether a brief global-access phase can learn a compatible expert-to-layer allocation that is then executed privately. UniPool-Lock gives every layer its own router over one global pool, balances aggregate pool usage, and scores experts with a scale-stable NormRouter. After the short ownership-learning phase, it assigns each expert to one layer, locks the disjoint allocation, and continues training as a layer-private MoE. With the same expert-FFN budget and routed expert FLOPs, this recipe lowers held-out loss relative to vanilla MoE by 0.024-0.037 across five dense-equivalent scales from 182M to 1.5B, including -0.0247 at 1.5B after 60B tokens. The ownership-learning phase lasts 2K steps, about 3.3% of training, and post-lock step time is within -0.7% to +2.8% of vanilla MoE in our throughput measurements. Controls locate the gain in full-pool training and in the allocation it produces: a random disjoint allocation fixed at initialization, trained with the same router and losses, matches vanilla, whereas locking the allocation learned during the full-pool phase recovers nearly all of the persistent full-pool gain, and substituting a random allocation at the lock forfeits about half of it. Keeping full-pool access throughout training (UniPool-Full) also improves over vanilla MoE at four scales and outperforms it with only 66.7% (182M) to 50% (469M and 650M) of its expert parameters.

cs.LG↗

arXiv2026

Less Uniform Discrete Diffusion is More Powerful and Scalable

Although uniform diffusion language models (UDLMs) represent a promising diffusion paradigm, scaling them remains challenging. We identify the core obstacle as an over-uniform training objective and condition-target confusion during sampling. To address these, we propose Less Uniform Diffusion (LUDI), a novel UDLM framework. Specifically, we (i) introduce a less uniform loss that directs each reverse transition toward the clean token, and (ii) equip the model with per-token time embeddings that supply token-level corruption hints, enabling confidence-based few-step sampling. Experiments across scales show that LUDI yields cleaner supervision and improves few-step generation. We further continue-train a 7B autoregressive model into LUDI-7B, resulting in a UDLM capable of complex reasoning. It achieves a 3-token-per-step speedup over AR decoding and competitive performance compared with masked diffusion baselines, revealing that the full potential of UDLMs for complex generation remains to be unlocked.

cs.CL↗

arXiv2026

AdaMTP: An Adaptive Training Paradigm for Multi-Token Prediction

Multi-Token Prediction (MTP) has emerged as an effective paradigm that augments a shared Large Language Model backbone with auxiliary heads, training the model to predict several future tokens in parallel to enrich its supervision signal and accelerate inference. However, existing training frameworks adopt a rigid, fixed-length prediction horizon, disregarding the highly non-uniform information density of natural language and code. Forcing the auxiliary heads to predict across high-entropy semantic boundaries injects noisy, conflicting training signals; because these heads share the backbone's latent representations, the resulting gradients backpropagate and interfere with the model's core capabilities. We propose AdaMTP, an adaptive training paradigm that dynamically aligns the prediction horizon with the intrinsic predictability of the sequence. At its core, an entropy-based segmentation algorithm leverages the base model to detect sudden surges in uncertainty as semantic boundaries, partitioning sequences into variable-length groups. Each token is assigned an adaptive prediction depth, and a dynamically masked MTP objective suppresses the loss for predictions that cross these boundaries, attenuating the noisy gradients that degrade the backbone. Across mathematical reasoning, code generation, and general benchmarks on three backbones (Llama-3.1-8B, Qwen-2.5-7B, Gemma-3-12B), AdaMTP consistently outperforms standard MTP in both task performance and inference speedup.

cs.CL↗

arXiv2026

Boogu-Image-0.1: Boosting Open Agentic Multimodal Generation via Understanding under a Minimal Budget

We introduce Boogu-Image-0.1, an open-source unified multimodal understanding and generation model family, comprising Base, Turbo, Edit, and Edit-Turbo variants. It delivers competitive performance in high-quality text-to-image generation, fast inference, instruction-based editing, and bilingual (Chinese-English) text rendering. Closed-source multimodal systems like Nano-Banana-Pro and GPT-Image-2 achieve strong performance through system-level integration rather than a single model, yet their internal practices remain largely undisclosed. In this work, we demonstrate that strengthening the understanding capability of the system, through a stronger multimodal encoder, agentic prompt rewriting, and related techniques, together with improvements in data quality, training pipelines, and agentic inference-time scaling, can substantially enhance generation and editing performance even under highly constrained compute budgets. Comprehensive evaluations show that Boogu-Image-0.1 consistently matches or surpasses other open-source models across standard benchmarks, and achieves results approaching leading closed-source systems. Notably, this is accomplished with only 208.62 million unique images. The base model's theoretical training cost is only approximately \$400K. We share practical discussions that we believe are valuable to the broader research community, and release weights, code, and recipes under Apache 2.0 to advance the open ecosystem for unified multimodal understanding and generation. Our code is available here: https://github.com/Boogu-Project/Boogu-Image.

cs.CV↗

arXiv2026

POET-X: Memory-efficient LLM Training by Scaling Orthogonal Transformation

Efficient and stable training of large language models (LLMs) remains a core challenge in modern machine learning systems. To address this challenge, Reparameterized Orthogonal Equivalence Training (POET), a spectrum-preserving framework that optimizes each weight matrix through orthogonal equivalence transformation, has been proposed. Although POET provides strong training stability, its original implementation incurs high memory consumption and computational overhead due to intensive matrix multiplications. To overcome these limitations, we introduce POET-X, a scalable and memory-efficient variant that performs orthogonal equivalence transformations with significantly reduced computational cost. POET-X maintains the generalization and stability benefits of POET while achieving substantial improvements in throughput and memory efficiency. In our experiments, POET-X enables the pretraining of billion-parameter LLMs on a single Nvidia H100 GPU, and in contrast, standard optimizers such as AdamW run out of memory under the same settings.

cs.LG↗

arXiv2026

ADRA-Bank: A Modular Benchmark for Academic Deep Research Agents

A surge in academic publications calls for automated deep research (DR) systems, but accurately evaluating them is still an open problem. First, existing benchmarks often focus narrowly on retrieval while neglecting high-level planning and reasoning. Second, existing benchmarks favor general domains over the academic domains that are the core application for DR agents. To address these gaps, we introduce ADRA-Bank, a modular benchmark for Academic DR Agents. Grounded in academic literature, our benchmark is a human-annotated dataset of 200 instances across 10 academic domains, including both research and review papers. Furthermore, we propose a modular Evaluation Paradigm for Academic DR Agents (ADRA-Eval), which leverages the rich structure of academic papers to assess the core capabilities of planning, retrieval, and reasoning. It employs two complementary modes: an end-to-end evaluation for \task agents and an isolated evaluation for foundational LLMs as potential backbones. Results reveal uneven capabilities: while agents show specialized strengths, they struggle with multi-source retrieval and cross-field consistency. Moreover, improving high-level planning capability is the crucial factor for unlocking the reasoning potential of foundational LLMs as backbones. By exposing these actionable failure modes, ADRA-Bank provides a diagnostic tool to guide the development of more reliable automatic academic research assistants.

cs.CL↗

arXiv2026

A Survey of Audio Reasoning in Multimodal Foundation Models

Reasoning has become a defining capability of modern foundation models, yet its development in the audio modality remains limited. Audio poses challenges that are distinct from those of text and vision. It is continuous, temporally dense, and contains linguistic, paralinguistic, and environmental information at multiple time scales. As a result, audio reasoning models must align acoustic signals with the discrete semantic space of large language models, while still preserving fine-grained information needed for reliable inference. Progress is also limited by three major obstacles: the scarcity of genuinely audio-grounded reasoning data, shortcut learning and modality hallucination, and the tension between reasoning depth and real-time latency in spoken interaction. In this paper, we present the first dedicated survey of audio reasoning. We provide a unified formulation that distinguishes direct predictive modeling from reasoning-augmented generation, review the architectural and training foundations of audio reasoning models, and systematically organize recent advances in Audio-to-Text, Audio-to-Speech, Audio-Visual Reasoning and Agentic Audio Reasoning. We further examine emerging paradigms such as Chain-of-Thought prompting, supervised fine-tuning, reinforcement learning, and latency-aware spoken interaction, and discuss evaluation practices, open challenges, and future directions. Our goal is to offer a coherent roadmap for developing robust, efficient, and natively grounded audio reasoning systems.

eess.AS↗

arXiv2026

FastCode: Fast and Cost-Efficient Code Understanding and Reasoning

Repository-scale code reasoning is a cornerstone of modern AI-assisted software engineering, enabling Large Language Models (LLMs) to handle complex workflows from program comprehension to complex debugging. However, balancing accuracy with context cost remains a significant bottleneck, as existing agentic approaches often waste computational resources through inefficient, iterative full-text exploration. To address this, we introduce FastCode, a framework that decouples repository exploration from content consumption. FastCode utilizes a structural scouting mechanism to navigate a lightweight semantic-structural map of the codebase, allowing the system to trace dependencies and pinpoint relevant targets without the overhead of full-text ingestion. By leveraging structure-aware navigation tools regulated by a cost-aware policy, the framework constructs high-value contexts in a single, optimized step. Extensive evaluations on the SWE-QA, LongCodeQA, LOC-BENCH, and GitTaskBench benchmarks demonstrate that FastCode consistently outperforms state-of-the-art baselines in reasoning accuracy while significantly reducing token consumption, validating the efficiency of scouting-first strategies for large-scale code reasoning. Source code is available at https://github.com/HKUDS/FastCode.

cs.SE↗

arXiv2026

Test-Time Learning of Causal Structure from Interventional Data

Supervised causal learning has shown promise in causal discovery, yet it often struggles with generalization across diverse interventional settings, particularly when intervention targets are unknown. To address this, we propose TICL (Test-time Interventional Causal Learning), a novel method that synergizes Test-Time Training with Joint Causal Inference. Specifically, we design a self-augmentation strategy to generate instance-specific training data at test time, effectively avoiding distribution shifts. Furthermore, by integrating joint causal inference, we developed a PC-inspired two-phase supervised learning scheme, which effectively leverages self-augmented training data while ensuring theoretical identifiability. Extensive experiments on bnlearn benchmarks demonstrate TICL's superiority in multiple aspects of causal discovery and intervention target detection.

cs.LG↗

arXiv2025

Survey and Experiments on Mental Disorder Detection via Social Media: From Large Language Models and RAG to Agents

Mental disorders represent a critical global health challenge, and social media is increasingly viewed as a vital resource for real-time digital phenotyping and intervention. To leverage this data, large language models (LLMs) have been introduced, offering stronger semantic understanding and reasoning than traditional deep learning, thereby enhancing the explainability of detection results. Despite the growing prominence of LLMs in this field, there is a scarcity of scholarly works that systematically synthesize how advanced enhancement techniques, specifically Retrieval-Augmented Generation (RAG) and Agentic systems, can be utilized to address these reliability and reasoning limitations. Here, we systematically survey the evolving landscape of LLM-based methods for social media mental disorder analysis, spanning standard pre-trained language models, RAG to mitigate hallucinations and contextual gaps, and agentic systems for autonomous reasoning and multi-step intervention. We organize existing work by technical paradigm and clinical target, extending beyond common internalizing disorders to include psychotic disorders and externalizing behaviors. Additionally, the paper comprehensively evaluates the performance of LLMs, including the impact of RAG, across various tasks. This work establishes a unified benchmark for the field, paving the way for the development of trustworthy, autonomous AI systems that can deliver precise and explainable mental health support.

cs.CL↗

arXiv2025

SJD++: Improved Speculative Jacobi Decoding for Training-free Acceleration of Discrete Auto-regressive Text-to-Image Generation

Large autoregressive models can generate high-quality, high-resolution images but suffer from slow generation speed, because these models require hundreds to thousands of sequential forward passes for next-token prediction during inference. To accelerate autoregressive text-to-image generation, we propose Speculative Jacobi Decoding++ (SJD++), a training-free probabilistic parallel decoding algorithm. Unlike traditional next-token prediction, SJD++ performs multi-token prediction in each forward pass, drastically reducing generation steps. Specifically, it integrates the iterative multi-token prediction mechanism from Jacobi decoding, with the probabilistic drafting-and-verification mechanism from speculative sampling. More importantly, for further acceleration, SJD++ reuses high-confidence draft tokens after each verification phase instead of resampling them all. We conduct extensive experiments on several representative autoregressive text-to-image generation models and demonstrate that SJD++ achieves $2\times$ to $3\times$ inference latency reduction and $2\times$ to $7\times$ step compression, while preserving visual quality with no observable degradation.

cs.CV↗

arXiv2025

Answer-Consistent Chain-of-thought Reinforcement Learning For Multi-modal Large Langauge Models

Recent advances in large language models (LLMs) have demonstrated that reinforcement learning with verifiable rewards (RLVR) can significantly enhance reasoning abilities by directly optimizing correctness, rather than relying solely on supervised imitation. This paradigm has been extended to multimodal LLMs for complex video and image understanding tasks. However, while outcome-driven RL improves answer accuracy, it can inadvertently decouple the reasoning chain from the final answer, leading to situations where models produce inconsistency between the reasoning trace and final answer. In our experiments on multiple-choice visual question-answering tasks, the standard GRPO method yields only 79.7\% consistency on MMVU between the reasoning steps and the chosen answers, indicating frequent mismatches between answers and reasoning. To this end, we propose Answer-Consistent Reinforcement Learning (ACRE) that modifies the GRPO algorithm with an auxiliary consistency check. After the model generates a chain of thought and an initial answer for a given question, we shuffle the answer options and prompt the model again with the same reasoning trace to predict a second answer. We design a consistency-verification reward that grants a high reward only if both the original and the post-shuffle answers agree and are correct; otherwise, a lower reward is assigned accordingly. This mechanism penalizes reasoning-answer misalignment and discourages the model from relying on spurious patterns, such as option ordering biases. We evaluate ACRE on challenging Video Reasoning benchmarks and multimodal math reasoning benchmarks, achieving an average 2.2\% and 1.5\% improvement for Video Reasoning and Math Reasoning tasks over the GRPO baseline.

cs.CV↗

arXiv2025

Speculative Jacobi-Denoising Decoding for Accelerating Autoregressive Text-to-image Generation

As a new paradigm of visual content generation, autoregressive text-to-image models suffer from slow inference due to their sequential token-by-token decoding process, often requiring thousands of model forward passes to generate a single image. To address this inefficiency, we propose Speculative Jacobi-Denoising Decoding (SJD2), a framework that incorporates the denoising process into Jacobi iterations to enable parallel token generation in autoregressive models. Our method introduces a next-clean-token prediction paradigm that enables the pre-trained autoregressive models to accept noise-perturbed token embeddings and predict the next clean tokens through low-cost fine-tuning. This denoising paradigm guides the model towards more stable Jacobi trajectories. During inference, our method initializes token sequences with Gaussian noise and performs iterative next-clean-token-prediction in the embedding space. We employ a probabilistic criterion to verify and accept multiple tokens in parallel, and refine the unaccepted tokens for the next iteration with the denoising trajectory. Experiments show that our method can accelerate generation by reducing model forward passes while maintaining the visual quality of generated images.

cs.CV↗

arXiv2025

SheetDesigner: MLLM-Powered Spreadsheet Layout Generation with Rule-Based and Vision-Based Reflection

Spreadsheets are critical to data-centric tasks, with rich, structured layouts that enable efficient information transmission. Given the time and expertise required for manual spreadsheet layout design, there is an urgent need for automated solutions. However, existing automated layout models are ill-suited to spreadsheets, as they often (1) treat components as axis-aligned rectangles with continuous coordinates, overlooking the inherently discrete, grid-based structure of spreadsheets; and (2) neglect interrelated semantics, such as data dependencies and contextual links, unique to spreadsheets. In this paper, we first formalize the spreadsheet layout generation task, supported by a seven-criterion evaluation protocol and a dataset of 3,326 spreadsheets. We then introduce SheetDesigner, a zero-shot and training-free framework using Multimodal Large Language Models (MLLMs) that combines rule and vision reflection for component placement and content population. SheetDesigner outperforms five baselines by at least 22.6\%. We further find that through vision modality, MLLMs handle overlap and balance well but struggle with alignment, necessitates hybrid rule and visual reflection strategies. Our codes and data is available at Github.

cs.AI↗

arXiv2025

Logits-Based Finetuning

In recent years, developing compact and efficient large language models (LLMs) has emerged as a thriving area of research. Traditional Supervised Fine-Tuning (SFT), which relies on singular ground truth labels, often fails to capture token-level dependencies and linguistic diversity. To address these limitations, we propose a logits-based fine-tuning framework that integrates the strengths of supervised learning and knowledge distillation. Our approach constructs enriched training targets by combining teacher logits with ground truth labels, preserving both correctness and linguistic diversity. This ensures more reliable and effective training. We constructed a large-scale 1.2M logits dataset and trained a series of science-focused models. Experimental results demonstrate that our method achieves significant improvements, with accuracy gains of 18% on Mawps and 22.7% on TabMWP. Across nine widely used mathematical benchmarks, our method consistently outperforms prior SFT models, achieving an average improvement of 7.28%. Codes are available at https://github.com/dvlab-research/Logits-Based-Finetuning.

cs.LG↗

arXiv2025

SepLLM: Accelerate Large Language Models by Compressing One Segment into One Separator

Large Language Models (LLMs) have exhibited exceptional performance across a spectrum of natural language processing tasks. However, their substantial sizes pose considerable challenges, particularly in computational demands and inference speed, due to their quadratic complexity. In this work, we have identified a key pattern: certain seemingly meaningless separator tokens (i.e., punctuations) contribute disproportionately to attention scores compared to semantically meaningful tokens. This observation suggests that information of the segments between these separator tokens can be effectively condensed into the separator tokens themselves without significant information loss. Guided by this insight, we introduce SepLLM, a plug-and-play framework that accelerates inference by compressing these segments and eliminating redundant tokens. Additionally, we implement efficient kernels for training acceleration. Experimental results across training-free, training-from-scratch, and post-training settings demonstrate SepLLM's effectiveness. Notably, using the Llama-3-8B backbone, SepLLM achieves over 50% reduction in KV cache on the GSM8K-CoT benchmark while maintaining comparable performance. Furthermore, in streaming settings, SepLLM effectively processes sequences of up to 4 million tokens or more while maintaining consistent language modeling capabilities.

cs.CL↗