arXiv subjects

Shichao Weng

Publications and source records attributed to Shichao Weng.

8 recordsLinked to original sources

arXiv2026

Meta-Learning Where to Allocate Experts: Task-Conditioned Layer-Wise Compression for MoEs

Mixture-of-Experts (MoE) models route each token to a subset of expert networks, increasing capacity while keeping per-token computation sparse. In many deployed MoEs, the number of active experts is fixed across layers and tasks, although layer roles and expert redundancy vary with depth and demand varies with difficulty. Existing approaches address only part of this setting: layer-wise allocations are usually determined offline and reused for all tasks, while token-level methods vary expert activation using local routing signals without task-level context. We propose MetaNet, a support-set controller that predicts, for each layer, an expert-retention threshold and a bounded routing bias. The backbone, experts, and router remain frozen. On DeepSeek-MoE-16B-Chat, MetaNet provides a tunable accuracy-expert-activation trade-off. Relative to fixed k=6, a conservative setting activates 3.61 experts on average (40% fewer) and achieves comparable MMLU accuracy (0.489 vs. 0.474), whereas an aggressive setting activates 2.28 experts on average (62% fewer) with accuracy approximately 3.7 percentage points lower. The MMLU-trained controller also transfers to C-Eval without retraining, activating 2.90 experts on average (52% fewer than fixed k=6) at 0.386 accuracy.

cs.CL↗

arXiv2026

BUDDY: BUdget-Driven DYnamic Depth Routing for Adaptive Large Language Model Inference

Large language models (LLMs) incur high inference cost due to their depth and parameter scale. Depth pruning can reduce latency by skipping redundant Transformer blocks, but existing methods (i) provide limited control under user-specific compute budgets and (ii) typically fix the routing path, failing to adapt as the context grows during decoding. We propose Buddy, a budget-driven dynamic depth routing framework. Buddy uses a lightweight Decision Module to score intermediate layers conditioned on the input and deterministically executes the top-k layers to satisfy a given budget. To support decode-time adaptation, Buddy reuses the first-layer KV cache as a low-overhead global context source and pools it together with the newest token representation before each routing decision. When no explicit budget is provided, an optional Budget Predictor estimates an input-dependent compute level to balance quality and efficiency. Experiments on Llama-family and Qwen models show that Buddy is competitive with strong static pruning baselines and often improves the accuracy-compute trade-off, while uniquely supporting strict budget control, decode-time rerouting, and multiple budgets within a single trained model.

cs.LG↗

arXiv2026

From 0-Order Selection to 2-Order Judgment: Combinatorial Hardening Exposes Compositional Failures in Frontier LLMs

Multiple-choice reasoning benchmarks face dual challenges: rapid saturation from advancing models and data contamination that undermines static evaluations. Ad-hoc hardening methods (paraphrasing, perturbation) attempt to increase difficulty but sacrifice logical validity for surface complexity, falling short to challenge advanced reasoning models. We present LogiHard, a formal framework that deterministically transforms 0-order selection into 2-order logical judgment, which significantly increases the thinking overhead and reasoning steps. The framework integrates Item Response Theory (IRT) for computerized adaptive testing (CAT), enabling precise difficulty control with fewer questions than static benchmarks. We instantiate LogiHard-2k, a logical reasoning dataset constructed by cognitively ranking high-stakes examination questions via 9-dimensional analysis of model thinking traces, followed by combinatorial transformation of high-difficulty items. Evaluation across twelve state-of-the-art models reveals an accuracy degradation ranging from 31% to 56% on combinatorially hardened questions. LLMs suffer from the multi-select failure and early exit bias, which are not shared by human testees. Zero-shot transfer to MMLU demonstrates 47% accuracy degradation (89.84% to 42.86%), confirming applicability across domains with provable validity preservation. The consistent aggregate degeneration is domain-agnostic and stems not from knowledge deficits but from a combinatorial reasoning gap, reflecting a training-induced completeness-verification deficit.

cs.CL↗

arXiv2026

CoRE: A Fine-Grained Code Reasoning Benchmark Beyond Output Prediction

Despite strong performance on code generation tasks, it remains unclear whether large language models (LLMs) genuinely reason about code execution. Existing code reasoning benchmarks primarily evaluate final output correctness under a single canonical implementation, leaving two critical aspects underexplored: (1) whether LLMs can maintain consistency to functionally equivalent implementations, and (2) whether LLMs can accurately reason about intermediate execution states. We introduce \textbf{CoRE}, a \textbf{Co}de \textbf{Re}asoning benchmark that evaluates code reasoning through \textbf{implementation invariance} and \textbf{process transparency}. Extensive evaluations on eight frontier LLMs reveal two fundamental limitations. First, models exhibit a substantial \textbf{robustness gap}, with performance varying significantly across equivalent implementations. Second, we observe \textbf{superficial execution}, where models arrive at correct final outputs without correctly reasoning about intermediate execution states. Together, these findings demonstrate that output-only evaluations are insufficient for assessing code reasoning and position CoRE as a necessary benchmark for evaluating robust and faithful code reasoning.\footnote{Data and code are available at https://github.com/ZJUSig/CoRE.}

cs.SE↗

arXiv2025

GeoSketch: A Neural-Symbolic Approach to Geometric Multimodal Reasoning with Auxiliary Line Construction and Affine Transformation

Geometric Problem Solving (GPS) poses a unique challenge for Multimodal Large Language Models (MLLMs), requiring not only the joint interpretation of text and diagrams but also iterative visuospatial reasoning. While existing approaches process diagrams as static images, they lack the capacity for dynamic manipulation - a core aspect of human geometric reasoning involving auxiliary line construction and affine transformations. We present GeoSketch, a neural-symbolic framework that recasts geometric reasoning as an interactive perception-reasoning-action loop. GeoSketch integrates: (1) a Perception module that abstracts diagrams into structured logic forms, (2) a Symbolic Reasoning module that applies geometric theorems to decide the next deductive step, and (3) a Sketch Action module that executes operations such as drawing auxiliary lines or applying transformations, thereby updating the diagram in a closed loop. To train this agent, we develop a two-stage pipeline: supervised fine-tuning on 2,000 symbolic-curated trajectories followed by reinforcement learning with dense, symbolic rewards to enhance robustness and strategic exploration. To evaluate this paradigm, we introduce the GeoSketch Benchmark, a high-quality set of 390 geometry problems requiring auxiliary construction or affine transformations. Experiments on strong MLLM baselines demonstrate that GeoSketch significantly improves stepwise reasoning accuracy and problem-solving success over static perception methods. By unifying hierarchical decision-making, executable visual actions, and symbolic verification, GeoSketch advances multimodal reasoning from static interpretation to dynamic, verifiable interaction, establishing a new foundation for solving complex visuospatial problems.

cs.AI↗

arXiv2025

Large Language Model Compression with Global Rank and Sparsity Optimization

Low-rank and sparse composite approximation is a natural idea to compress Large Language Models (LLMs). However, such an idea faces two primary challenges that adversely affect the performance of existing methods. The first challenge relates to the interaction and cooperation between low-rank and sparse matrices, while the second involves determining weight allocation across different layers, as redundancy varies considerably among them. To address these challenges, we propose a novel two-stage LLM compression method with the capability of global resource allocation for rank and sparsity. It is noteworthy that the overall optimization space is vast, making comprehensive optimization computationally prohibitive. Therefore, to reduce the optimization space, our first stage utilizes robust principal component analysis to decompose the weight matrices of LLMs into low-rank and sparse components, which span the low dimensional and sparse spaces containing the resultant low-rank and sparse matrices, respectively. In the second stage, we propose a probabilistic global allocation strategy to jointly identify the low-rank and sparse structures within the above two spaces. The appealing feature of our approach is its ability to automatically detect the redundancy across different layers and to manage the interaction between the sparse and low-rank components. Extensive experimental results indicate that our method significantly surpasses state-of-the-art techniques for sparsification and composite approximation.

cs.LG↗

arXiv2025

Balancing Fidelity and Plasticity: Aligning Mixed-Precision Fine-Tuning with Linguistic Hierarchies

Deploying and fine-tuning Large Language Models (LLMs) on resource-constrained edge devices requires navigating a strict trade-off between memory footprint and task performance. While Quantization-Aware Fine-tuning has emerged as a viable solution, existing paradigms typically decouple quantization and adapter optimization. This separation overlooks a fundamental theoretical constraint we identify as the \textit{Fidelity-Plasticity Trade-off}: a layer's capacity to adapt to new tasks (Plasticity) is inherently constrained by the information capacity of its frozen weights (Fidelity). Aggressively quantizing semantically critical layers creates an information bottleneck that no amount of adapter rank can recover, while high precision in robust syntactic layers wastes valuable memory. To address this, we introduce \textbf{QR-Adaptor}, a unified framework that jointly optimizes per-layer quantization bit-width and LoRA rank. By formulating resource allocation as a multi-objective search aligned with the model's linguistic hierarchy, our method systematically liberates memory from redundancy-heavy layers to reinvest in capacity-critical ones. Extensive experiments demonstrate that QR-Adaptor establishes a new Pareto frontier: notably, a model fine-tuned under a strict 4-bit memory budget achieves performance rivaling 16-bit baselines, demonstrating that precise resource alignment is as critical as model size.

cs.LG↗

arXiv2024

AutoMixQ: Self-Adjusting Quantization for High Performance Memory-Efficient Fine-Tuning

Fine-tuning large language models (LLMs) under resource constraints is a significant challenge in deep learning. Low-Rank Adaptation (LoRA), pruning, and quantization are all effective methods for improving resource efficiency. However, combining them directly often results in suboptimal performance, especially with uniform quantization across all model layers. This is due to the complex, uneven interlayer relationships introduced by pruning, necessitating more refined quantization strategies. To address this, we propose AutoMixQ, an end-to-end optimization framework that selects optimal quantization configurations for each LLM layer. AutoMixQ leverages lightweight performance models to guide the selection process, significantly reducing time and computational resources compared to exhaustive search methods. By incorporating Pareto optimality, AutoMixQ balances memory usage and performance, approaching the upper bounds of model capability under strict resource constraints. Our experiments on widely used benchmarks show that AutoMixQ reduces memory consumption while achieving superior performance. For example, at a 30\% pruning rate in LLaMA-7B, AutoMixQ achieved 66.21\% on BoolQ compared to 62.45\% for LoRA and 58.96\% for LoftQ, while reducing memory consumption by 35.5\% compared to LoRA and 27.5\% compared to LoftQ.

cs.LG↗