arXiv subjects

Anhao Zhao

Publications and source records attributed to Anhao Zhao.

At least 19 recordsLinked to original sources

arXiv2026

Post-Training Shifts Confidence: A Three-Stage Analysis of How SFT, RL, and OPD Shape CoT Calibration

Large language models have made strong reasoning gains through supervised fine-tuning, reinforcement learning, and on-policy distillation, yet these post-training methods are usually evaluated only by final-answer accuracy. We study how they reshape confidence during reasoning. We introduce a three-stage calibration framework that evaluates confidence before, during, and after chain-of-thought generation, corresponding to difficulty estimation, early termination, and answer aggregation. Through a controlled comparison on mathematical reasoning benchmarks, we find that OPD provides the most useful pre-reasoning confidence, SFT gives the strongest online signal for early stopping, and RL produces the most reliable trace-level signal for aggregation. We further show that confidence reliability is position-dependent: RL confidence becomes informative after a path-commitment phase, while OPD confidence is useful early but can become inversely calibrated later. Based on this observation, we propose PosConf, a position-aware confidence strategy that uses confidence only from reliable relative-position intervals. PosConf improves RL answer aggregation by 6.1 points over majority voting and consistently improves OPD early stopping under tight token budgets, with gains up to 4.3 points by avoiding its later inverse-calibration region, showing that \emph{confidence in reasoning models should be used both stage-wise and position-awarely}. Our code is available at https://github.com/EIT-NLP/Post-Training-Calibration.

cs.CL↗

arXiv2026

PowerOPD: Stabilizing On-Policy Distillation with Bounded Power Transformation

Standard on-policy distillation (OPD) for large language models estimates the reverse-KL objective using student-sampled tokens, yielding an unbiased single-sample Monte Carlo estimator that avoids vocabulary-wide computation. However, we show that this estimator suffers from severe training pathologies in practice: sample inefficiency, unstable generation dynamics, and a substantial performance gap compared to exact full-vocabulary OPD. Reward-level diagnosis traces these pathologies to the log-ratio reward, which is unbounded by construction, producing extremely high-variance gradients concentrated at early positions and persisting throughout training; standard post-hoc scaling fail as they operate only after this distortion occurs. To solve this problem, we propose PowerOPD: a family of natively bounded, sign-consistent rewards from the Box-Cox power transformation, parameterized by alpha > 0, of which the log-ratio is the degenerate alpha -> 0 limit. Across six mathematical reasoning benchmarks and four Qwen3 teacher-student pairs, PowerOPD achieves benchmark-averaged Avg@8/Pass@8 gains of up to +6.37/+5.71 over vanilla OPD, +3.01/+3.54 over post-hoc stabilization, and +2.59/+8.90 over full-vocabulary OPD, while reducing wall-clock time by 59.2% and peak GPU memory by 23.1%. Larger alpha generally improves accuracy, consistently shortens responses, and keeps gradient norms more than 3,000x smaller than vanilla OPD.

cs.LG↗

arXiv2026

AdaSR: Adaptive Streaming Reasoning with Hierarchical Relative Policy Optimization

Large reasoning models typically follow a read-then-think paradigm: they observe the complete input, reason over a static context, and then produce the answer. Yet many real-world scenarios are inherently dynamic, such as audio and video stream, where information arrives as a continuous stream and models must reason, update, and respond under partial observations. Recent streaming reasoning methods allow models to think while reading, but they largely rely on supervised imitation of pre-constructed trajectories, which limits their flexibility. In this paper, we propose AdaSR, an adaptive streaming reasoning framework that enables models to reason during input streaming and perform final deliberation once the stream is complete, learning when to think, and how much computation to allocate across different stages. To optimize this hierarchical reasoning process, we introduce Hierarchical Relative Policy Optimization (HRPO), which decomposes policy optimization into streaming reasoning and deep reasoning phases, providing more fine-grained advantage assignment instead of uniformly distributing a single sequence-level advantage over all tokens. HRPO integrates format, accuracy, and adaptive thinking rewards to enforce valid reasoning protocols, preserve final task performance, and encourage latency-aware computation allocation. Experiments show that AdaSR achieves a better balance among reasoning accuracy, computational efficiency, and streaming latency compared with supervised fine-tuning baseline. We release our code at https://github.com/EIT-NLP/StreamingLLM/tree/main/AdaSR.

cs.CL↗

arXiv2026

miniReranker: Efficient Multimodal Reranking through Visual Cache Reuse and Interaction Sparsity

Multimodal large language models (MLLMs) have recently shown strong potential as point-wise rerankers by directly modeling query--document relevance through next-token prediction. However, point-wise reranking suffers from substantial repeated computation across query--document pairs, while the causal structure of transformers allows only prefix segments to be reused via pre-caching. To address the misalignment of existing query-first and document-first formats with both VQA-style prompting and computation-aware reuse, we propose a $\textit{vision-first}$ formulation that improves both cache reuse efficiency and reranking performance. However, the remaining cost is still considerable and stems from three main sources: (1) $\textit{model depth}$, for which we reduce active parameters via early exit; (2) $\textit{cross-segment attention}$, which we restrict to a narrow interaction band across a few layers; and (3) $\textit{visual tokens}$, where we reduce the number of tokens via embedder-guided pruning. Together, these designs form miniReranker, which reduces reranking runtime to <1% of the dense implementation under high-reuse settings for a single query, while preserving >96% of the dense model performance.

cs.IR↗

arXiv2026

Beyond FLOPs: Benchmarking Real Inference Acceleration of LLM Pruning under a GEMM-Centric Taxonomy

Pruning has emerged as a dominant paradigm for accelerating large language model (LLM) inference, spanning a broad spectrum of methods that remove computation across tokens, layers, heads, dimensions, and attention patterns. Despite sharing the same objective, these pruning approaches induce fundamentally different execution behaviors, causing realized speedups to depend heavily on hardware and kernel implementations. Consequently, the practical acceleration benefits of different pruning families remain poorly understood. In this work, we introduce a GEMM-centric taxonomy that reorganizes existing pruning methods according to the logical \textbf{M}, \textbf{N}, and \textbf{K} dimensions of general matrix multiplication (GEMM). Leveraging this abstraction, we build a unified benchmarking framework that enables implementation-consistent comparison across the pruning design space and systematically characterizes the acceleration--quality Pareto frontier. Our results on Llama3.1-8B show that static depth pruning remains the strongest Pareto-optimal baseline and stays closest to its theoretical acceleration upper bound in memory-bounded scenarios. During prefill, the frontier transitions from static depth at low quality loss (0\%--4\%), to dynamic depth at moderate loss (5\%--16\%), and finally to static width pruning at higher loss levels (17\%--26\%). These findings establish the first unified view of the practical limits of pruning-based LLM acceleration and provide guidance for future pruning research. Code is available at https://github.com/EIT-NLP/LLM-Pruning/tree/main/PruningInferSim

cs.LG↗

arXiv2026

Escaping the KL Agreement Trap in On-Policy Distillation

On-policy distillation (OPD) provides dense token-level supervision by asking a teacher to score student-generated rollouts. However, when the student drifts into an unrecoverable prefix, the teacher may locally agree with the degraded state, producing low reverse KL but little corrective training signal. We identify this persistent regime as a low-KL agreement trap. Further analyses show that tokens during and after such traps produce less useful supervision signals. We propose KAT (KL Agreement Trap Termination), an online OPD termination rule that detects persistent low-KL agreement with a dynamic training-adaptive threshold. By filtering weak supervision from degenerate agreement, KAT improves avg@k accuracy by 2.66% and pass@k by 3.43% across four mathematical benchmarks, while reducing average rollout length by 59.73%.

cs.LG↗

arXiv2026

ProactiveLLM: Learning Active Interaction for Streaming Large Language Models

Standard Large Language Models (LLMs) follow a read-then-generate paradigm, causing unnecessary latency and computation. Streaming LLMs alleviate this issue by generating while receiving inputs, but still struggle to decide when to interact with the stream. Existing methods either hard-code interaction timing or rely on costly external alignment signals, such as timing labels, reasoning trajectories, or stronger teachers. In this paper, we propose ProactiveLLM, which achieves active interaction by leveraging the model's endogenous states to guide interaction decisions. The model first learns to perceive semantic sufficiency from partial inputs through two complementary training mechanisms: mask-based streaming modeling and synchronized privileged self-distillation (SPSD). The former applies monotonic random masking to the input during training, simulating progressively revealed streaming inputs and enabling the model to learn local semantic dependencies from partial-input views. The latter aligns the partial-context student view with a full-context teacher view generated by the same evolving model, allowing privileged full-context evidence to guide the student's understanding under incomplete observations. Together, these mechanisms induce endogenous sufficiency cues without requiring external teachers or annotations, providing a versatile foundation for the plug-and-play integration of diverse decision heads. Extensive evaluation across text and speech streaming tasks confirms that ProactiveLLM significantly reduces interaction latency while maintaining quality, validating its capacity for dynamic and active interaction. Code is publicly available at https://github.com/EIT-NLP/StreamingLLM/tree/main/ProactiveLLM.

cs.CL↗

arXiv2026

Decoupling KL and Trajectories: A Unified Perspective for SFT, DAgger, Offline RL, and OPD in LLM Distillation

Knowledge distillation is central to LLM post-training, yet its design space remains poorly understood, especially alongside reinforcement learning (RL). We show that the prevailing paradigms, off-policy distillation and on-policy distillation (OPD), implicitly couple two orthogonal choices: prefix source and token-level KL direction. This follows from decomposing sequence-level KL over autoregressive response distributions: forward KL pairs teacher prefixes with token-level forward KL, and reverse KL pairs student prefixes with token-level reverse KL. We argue this coupling is not intrinsic: decoupling the two axes yields four valid objectives. We establish gradient-level identities showing forward KL gives SFT-style cross-entropy matching with teacher soft targets, whereas reverse KL gives an RL-style policy-gradient objective with a dense teacher-student log-ratio reward, connecting them to off-policy SFT, DAgger-style on-policy SFT, offline-RL-style distillation, and OPD. We conduct an extensive controlled study on math reasoning, evaluating the four objectives both as standalone methods and as initializations for subsequent RL. The results reveal three tradeoffs: KL direction induces an accuracy-entropy tradeoff, prefix source a quality-compute tradeoff, and training length an accuracy-stability tradeoff. Motivated by these findings, we propose KL mixing and an entropy-gated length curriculum. KL mixing shows long-sequence distillation requires substantial forward-KL weight to prevent entropy collapse and length inflation without sacrificing accuracy. The entropy-gated length curriculum improves Avg@k and Pass@k by 3.6 and up to 5.8 points, and cuts average response length by roughly 3x versus fixed long-horizon training. Our results provide a framework and practical methods for designing reasoning distillation objectives that balance accuracy, diversity, compute, and RL behavior.

cs.LG↗

arXiv2026

SkipOPU: An FPGA-based Overlay Processor for Large Language Models with Dynamically Allocated Computation

Large language models (LLMs) have achieved remarkable performance across a wide range of tasks, but their inference efficiency remains a critical bottleneck due to rapidly growing parameters. Recent advances in dynamic computation allocation address this challenge by exploiting the highly uneven contributions of different tokens and layers, enabling selective execution that significantly reduces redundant computation while preserving model accuracy. However, existing hardware platforms and accelerators are primarily optimized for uniform, static execution, limiting their ability to efficiently support such dynamic inference patterns. In this work, we propose SkipOPU, an FPGA-based overlay processor that dynamically allocates computation across tokens and layers with high flexibility through a lightweight routing mechanism. First, we decouple reduction operations from element-wise computation in nonlinear modules and perform reductions incrementally, which enables both stages to be fused with adjacent linear operations (router or matrix multiplication) for effective latency hiding. Second, motivated by asymmetric sensitivity to numerical precision between activation and weight, we design a PE array that efficiently supports float-fixed hybrid execution. A novel DSP overpacking technique is introduced to maximize hardware utilization while minimizing resource overhead. Finally, we develop a proactive on-chip KV history buffer that exploits cross-layer KV invariance of pruned tokens, eliminating irregular HBM accesses during decoding and supplementing off-chip bandwidth through high-locality on-chip reuse. Experimental results demonstrate that SkipOPU on an AMD U280 FPGA outperforms GPU and other FPGA-based accelerators by 1.23x-3.83x in bandwidth efficiency for LLMs inference with dynamic computation allocation and can reduce up to 25.4% KV storage overhead across varying sequence lengths.

cs.AR↗

arXiv2026

What Do Visual Tokens Really Encode? Uncovering Sparsity and Redundancy in Multimodal Large Language Models

Multimodal large language models (MLLMs) project visual tokens into the embedding space of language models, yet the internal structuring and processing of visual semantics remain poorly understood. In this work, we introduce a two-fold analytical framework featuring a novel probing tool, $\textbf{EmbedLens}$, to conduct a fine-grained analysis. We uncover a pronounced semantic sparsity at the input level: visual tokens consistently partition into sink, dead, and alive categories. Remarkably, only the alive tokens, comprising $\approx60\%$ of the total input, carry image-specific meaning. Furthermore, using a targeted patch-compression benchmark, we demonstrate that these alive tokens already encode rich, fine-grained cues (e.g., objects, colors, and OCR) prior to entering the LLM. Internal visual computations (such as visual attention and feed-forward networks) are redundant for most standard tasks. For the small subset of highly vision-centric tasks that actually benefit from internal processing, we reveal that alive tokens naturally align with intermediate LLM layers rather than the initial embedding space, indicating that shallow-layer processing is unnecessary and that direct mid-layer injection is both sufficient. Ultimately, our findings provide a unified mechanistic view of visual token processing, paving the way for more efficient and interpretable MLLM architectures through selective token pruning, minimized visual computation, and mid-layer injection. The code is released at: https://github.com/EIT-NLP/EmbedLens.

cs.CV↗

arXiv2026

On-Policy Supervised Fine-Tuning for Efficient Reasoning

Large reasoning models (LRMs) are commonly trained with reinforcement learning (RL) to explore long chain-of-thought reasoning, achieving strong performance at high computational cost. Recent methods add multi-reward objectives to jointly optimize correctness and brevity, but these complex extensions often destabilize training and yield suboptimal trade-offs. We revisit this objective and challenge the necessity of such complexity. Through principled analysis, we identify fundamental misalignments in this paradigm: KL regularization loses its intended role when correctness and length are directly verifiable, and group-wise normalization becomes ambiguous under multiple reward signals. By removing these two items and simplifying the reward to a truncation-based length penalty, we show that the optimization problem reduces to supervised fine-tuning on self-generated data filtered for both correctness and conciseness. We term this simplified training strategy on-policy SFT. Despite its simplicity, on-policy SFT consistently defines the accuracy-efficiency Pareto frontier. It reduces CoT length by up to 80 while maintaining original accuracy, surpassing more complex RL-based methods across five benchmarks. Furthermore, it significantly enhances training efficiency, reducing GPU memory usage by 50% and accelerating convergence by 70%. Our code is available at https://github.com/EIT-NLP/On-Policy-SFT.

cs.AI↗

arXiv2026

ViCA: Efficient Multimodal LLMs with Vision-Only Cross-Attention

Modern multimodal large language models (MLLMs) adopt a unified self-attention design that processes visual and textual tokens at every Transformer layer, incurring substantial computational overhead. In this work, we revisit the necessity of such dense visual processing and show that projected visual embeddings are already well-aligned with the language space, while effective vision-language interaction occurs in only a small subset of layers. Based on these insights, we propose ViCA (Vision-only Cross-Attention), a minimal MLLM architecture in which visual tokens bypass all self-attention and feed-forward layers, interacting with text solely through sparse cross-attention at selected layers. Extensive evaluations across three MLLM backbones, nine multimodal benchmarks, and 26 pruning-based baselines show that ViCA preserves 98% of baseline accuracy while reducing visual-side computation to 4%, consistently achieving superior performance-efficiency trade-offs. Moreover, ViCA provides a regular, hardware-friendly inference pipeline that yields over 3.5x speedup in single-batch inference and over 10x speedup in multi-batch inference, reducing visual grounding to near-zero overhead compared with text-only LLMs. It is also orthogonal to token pruning methods and can be seamlessly combined for further efficiency gains. Our code is available at https://github.com/EIT-NLP/ViCA.

cs.CV↗

arXiv2026

From LLMs to LRMs: Rethinking Pruning for Reasoning-Centric Models

Large language models (LLMs) are increasingly costly to deploy, motivating extensive research on model pruning. However, most existing studies focus on instruction-following LLMs, leaving it unclear whether established pruning strategies transfer to reasoning-augmented models that explicitly generate long intermediate reasoning traces. In this work, we conduct a controlled study of pruning for both instruction-following ($\textbf{LLM-instruct}$) and reasoning-augmented ($\textbf{LLM-think}$) models. To isolate the effects of pruning, we align pruning calibration and post-pruning recovery data with each model's original training distribution, which we show yields more stable and reliable pruning behavior. We evaluate static depth pruning, static width pruning, and dynamic pruning across 17 tasks spanning classification, generation, and reasoning. Our results reveal clear paradigm-dependent differences: depth pruning outperforms width pruning on classification tasks, while width pruning is more robust for generation and reasoning. Moreover, static pruning better preserves reasoning performance, whereas dynamic pruning excels on classification and generation but remains challenging for long-chain reasoning. These findings underscore the need for pruning strategies that explicitly account for the distinct characteristics of reasoning-augmented LLMs. Our code is publicly available at https://github.com/EIT-NLP/LRM-Pruning.

cs.LG↗

arXiv2025

$\mathcal{V}isi\mathcal{P}runer$: Decoding Discontinuous Cross-Modal Dynamics for Efficient Multimodal LLMs

Multimodal Large Language Models (MLLMs) have achieved strong performance across vision-language tasks, but suffer from significant computational overhead due to the quadratic growth of attention computations with the number of multimodal tokens. Though efforts have been made to prune tokens in MLLMs, \textit{they lack a fundamental understanding of how MLLMs process and fuse multimodal information.} Through systematic analysis, we uncover a \textbf{three-stage} cross-modal interaction process: (1) Shallow layers recognize task intent, with visual tokens acting as passive attention sinks; (2) Cross-modal fusion occurs abruptly in middle layers, driven by a few critical visual tokens; (3) Deep layers discard vision tokens, focusing solely on linguistic refinement. Based on these findings, we propose \emph{VisiPruner}, a training-free pruning framework that reduces up to 99\% of vision-related attention computations and 53.9\% of FLOPs on LLaVA-v1.5 7B. It significantly outperforms existing token pruning methods and generalizes across diverse MLLMs. Beyond pruning, our insights further provide actionable guidelines for training efficient MLLMs by aligning model architecture with its intrinsic layer-wise processing dynamics. Our code is available at: https://github.com/EIT-NLP/VisiPruner.

cs.CV↗

arXiv2025

StreamingThinker: Large Language Models Can Think While Reading

Large language models (LLMs) have demonstrated remarkable capabilities in chain of thought (CoT) reasoning. However, the current LLM reasoning paradigm initiates thinking only after the entire input is available, which introduces unnecessary latency and weakens attention to earlier information in dynamic scenarios. Inspired by human cognition of thinking while reading, we first design a \textit{\textbf{streaming thinking}} paradigm for LLMs, where reasoning unfolds in the order of input and further adjusts its depth once reading is complete. We instantiate this paradigm with \textit{StreamingThinker}, a framework that enables LLMs to think while reading through the integration of streaming CoT generation, streaming-constraint training, and streaming parallel inference. Specifically, StreamingThinker employs streaming reasoning units with quality control for CoT generation, enforces order-preserving reasoning through streaming attention masks and position encoding, and leverages parallel KV caches that decouple input encoding from reasoning generation, thereby ensuring alignment and enabling true concurrency. We evaluate StreamingThinker on the Qwen3 model family across math reasoning, logical reasoning, and context-based QA reasoning tasks. Experimental results show that the StreamingThinker preserves performance comparable to batch thinking, while yielding an 80\% reduction in token waiting before the onset of reasoning and a more than 60\% reduction in time-level latency for producing the final answer, demonstrating the effectiveness of the streaming paradigm for LLM reasoning. Code is publicly available at https://github.com/EIT-NLP/StreamingLLM/tree/main/StreamingThinker.

cs.CL↗

arXiv2025

SkipGPT: Dynamic Layer Pruning Reinvented with Token Awareness and Module Decoupling

Large language models (LLMs) achieve remarkable performance across tasks but incur substantial computational costs due to their deep, multi-layered architectures. Layer pruning has emerged as a strategy to alleviate these inefficiencies, but conventional static pruning methods overlook two critical dynamics inherent to LLM inference: (1) horizontal dynamics, where token-level heterogeneity demands context-aware pruning decisions, and (2) vertical dynamics, where the distinct functional roles of MLP and self-attention layers necessitate component-specific pruning policies. We introduce SkipGPT, a dynamic layer pruning framework designed to optimize computational resource allocation through two core innovations: (1) global token-aware routing to prioritize critical tokens, and (2) decoupled pruning policies for MLP and self-attention components. To mitigate training instability, we propose a two-stage optimization paradigm: first, a disentangled training phase that learns routing strategies via soft parameterization to avoid premature pruning decisions, followed by parameter-efficient LoRA fine-tuning to restore performance impacted by layer removal. Extensive experiments demonstrate that SkipGPT reduces over 40% of model parameters while matching or exceeding the performance of the original dense model across benchmarks. By harmonizing dynamic efficiency with preserved expressivity, SkipGPT advances the practical deployment of scalable, resource-aware LLMs. Our code is publicly available at: https://github.com/EIT-NLP/SkipGPT.

cs.CL↗

arXiv2025

Reasoning Beyond Language: A Comprehensive Survey on Latent Chain-of-Thought Reasoning

Large Language Models (LLMs) have shown impressive performance on complex tasks through Chain-of-Thought (CoT) reasoning. However, conventional CoT relies on explicitly verbalized intermediate steps, which constrains its broader applicability, particularly in abstract reasoning tasks beyond language. To address this, there has been growing research interest in \textit{latent CoT reasoning}, where the reasoning process is embedded within latent spaces. By decoupling reasoning from explicit language generation, latent CoT offers the promise of richer cognitive representations and facilitates more flexible, faster inference. This paper aims to present a comprehensive overview of this emerging paradigm and establish a systematic taxonomy. We analyze recent advances in methods, categorizing them from token-wise horizontal approaches to layer-wise vertical strategies. We then provide in-depth discussions of these methods, highlighting their design principles, applications, and remaining challenges. We hope that our survey provides a structured foundation for advancing this promising direction in LLM reasoning. The relevant papers will be regularly updated at https://github.com/EIT-NLP/Awesome-Latent-CoT.

cs.CL↗

arXiv2025

LLM as Effective Streaming Processor: Bridging Streaming-Batch Mismatches with Group Position Encoding

Large Language Models (LLMs) are primarily designed for batch processing. Existing methods for adapting LLMs to streaming rely either on expensive re-encoding or specialized architectures with limited scalability. This work identifies three key mismatches in adapting batch-oriented LLMs to streaming: (1) input-attention, (2) output-attention, and (3) position-ID mismatches. While it is commonly assumed that the latter two mismatches require frequent re-encoding, our analysis reveals that only the input-attention mismatch significantly impacts performance, indicating re-encoding outputs is largely unnecessary. To better understand this discrepancy with the common assumption, we provide the first comprehensive analysis of the impact of position encoding on LLMs in streaming, showing that preserving relative positions within source and target contexts is more critical than maintaining absolute order. Motivated by the above analysis, we introduce a group position encoding paradigm built on batch architectures to enhance consistency between streaming and batch modes. Extensive experiments on cross-lingual and cross-modal tasks demonstrate that our method outperforms existing approaches. Our method requires no architectural modifications, exhibits strong generalization in both streaming and batch modes. The code is available at repository https://github.com/EIT-NLP/StreamingLLM.

cs.CL↗