← Reading List
📰 技术速递

技术速递 2026-03-24:LLM 推理评估与能力边界

Tech多源前沿LLM推理评估arXiv

每日技术速递 · 2026-03-24

今日聚焦:LLM 对话推理能力、Chain-of-Thought 忠实度评估、语义 Token 聚类、视觉 Token 剪枝等前沿方向。


1. Reasoning Gets Harder for LLMs Inside A Dialogue

来源: arXiv cs.CL 链接: https://arxiv.org/abs/2603.20133v1

Large Language Models (LLMs) achieve strong performance on many reasoning benchmarks, yet these evaluations typically focus on isolated tasks that differ from real-world usage in task-oriented dialogue (TOD). In this setting, LLMs must perform reason…


2. LoASR-Bench: Evaluating Large Speech Language Models on Low-Resource Automatic Speech Recognition Across Language Families

来源: arXiv cs.CL 链接: https://arxiv.org/abs/2603.20042v1

Large language models (LLMs) have driven substantial advances in speech language models (SpeechLMs), yielding strong performance in automatic speech recognition (ASR) under high-resource conditions. However, existing benchmarks predominantly focus on…


3. Measuring Faithfulness Depends on How You Measure: Classifier Sensitivity in LLM Chain-of-Thought Evaluation

来源: arXiv cs.LG 链接: https://arxiv.org/abs/2603.20172v1

Recent work on chain-of-thought (CoT) faithfulness reports single aggregate numbers (e.g., DeepSeek-R1 acknowledges hints 39% of the time), implying that faithfulness is an objective, measurable property of a model. This paper demonstrates that it is…


4. Semantic Token Clustering for Efficient Uncertainty Quantification in Large Language Models

来源: arXiv cs.LG 链接: https://arxiv.org/abs/2603.20161v1

Large language models (LLMs) have demonstrated remarkable capabilities across diverse tasks. However, the truthfulness of their outputs is not guaranteed, and their tendency toward overconfidence further limits reliability. Uncertainty quantification…


5. Current LLMs still cannot ‘talk much’ about grammar modules: Evidence from syntax

来源: arXiv cs.CL 链接: https://arxiv.org/abs/2603.20114v1

We aim to examine the extent to which Large Language Models (LLMs) can ‘talk much’ about grammar modules, providing evidence from syntax core properties translated by ChatGPT into Arabic. We collected 44 terms from generative syntax previous works, i…


6. Frequency Switching Mechanism for Parameter-E!cient Multi-Task Learning

来源: arXiv cs.LG (systems) 链接: https://arxiv.org/abs/2603.21111v1

Multi-task learning (MTL) aims to enable a single model to solve multiple tasks efficiently; however, current parameter-efficient fine-tuning (PEFT) methods remain largely limited to single-task adaptation. We introduce \textbf{Free Sinewich}, a para…


7. ResPrune: Text-Conditioned Subspace Reconstruction for Visual Token Pruning in Large Vision-Language Models

来源: arXiv cs.LG (systems) 链接: https://arxiv.org/abs/2603.21105v1

Large Vision-Language Models (LVLMs) rely on dense visual tokens to capture fine-grained visual information, but processing all these tokens incurs substantial computational and memory overhead during inference. To address this issue, we propose ResP…


*🔗 查看 iWiki 完整版 自动生成于 2026-03-24*

评论