计算语言学：2025–2026 年度综述

An annual survey of the cs.CL arXiv category, covering research published from September 2025 – August 2026. Built on a close-reading set of 144 papers and organised into 40 chapters, it maps the year by the field's research sub-directions.

Book framework (share of the year's output):

• LLM Reasoning, Benchmarking, and Multimodal Evaluation (45.1%) — Focuses on evaluating and improving large language models' reasoning capabilities, multimodal understanding, and performance on complex benchmarks.
• Vision-Language Models and Robotic Skill Adaptation (15.0%) — Explores the capabilities of vision-language models in spatial reasoning and visual math, alongside frameworks for robot skill learning and adaptation.
• Speech Processing and Dialogue System Assessment (11.5%) — Addresses challenges in automated speaking assessment, neural TTS modeling, and proactive dialogue systems for human-computer interaction.
• Retrieval-Augmented Generation and Multi-Agent Knowledge Systems (7.1%) — Investigates RAG architectures, multi-agent collaboration, and knowledge graph integration for generating accurate, evidence-based answers in specialized domains.
• Adaptive Retrieval and Large-Scale Corpus Processing (4.2%) — Develops hierarchical and adaptive retrieval techniques for handling complex queries and large-scale corpora across diverse languages and domains.
• Reinforcement Learning and Agentic Optimization for LLMs (3.9%) — Applies reinforcement learning and preference optimization to enhance LLM consistency, topic modeling, and agentic decision-making policies.
• Semantic Representation and Uncertainty Quantification (3.9%) — Examines latent space representations, semantic navigation, and methods for quantifying uncertainty in long-form language model outputs.
• Machine-Generated Text Detection and Content Classification (3.5%) — Focuses on detecting AI-generated text and code, auditing generative content, and improving classification models for web and legal content.
• Multilingual Adaptation and Parameter-Efficient Fine-Tuning (2.7%) — Enhances multilingual LLM performance through low-resource adaptation, mixture-of-experts routing, and parameter-efficient fine-tuning techniques.
• Synthetic Data Generation and Augmentation Strategies (0.7%) — Explores methods for synthesizing high-quality training data, including text, audio, and code, to improve model training efficiency and robustness.
• Adversarial Attacks and Prompt Injection Robustness (0.7%) — Analyzes jailbreak attacks, adversarial suffixes, and prompt injection vulnerabilities to improve the robustness of LLMs against malicious inputs.
• Safety Alignment and Selective Unlearning (0.5%) — Addresses LLM safety alignment, tokenization bias, and selective unlearning techniques to mitigate risks and ensure responsible AI behavior.

Each sub-direction is covered across three chapters — how it framed its questions, the strength of its key findings, and its open debates and frontiers — so the reader gets a traceable picture of what this part of the field actually produced over the year.