Unpacking the ‘Thought’ in AI: Recent Advances in Chain-of-Thought Reasoning for LLMs and MLLMs

Latest 50 papers on chain-of-thought reasoning: Nov. 2, 2025

The ability of AI models to “think” step-by-step, often referred to as Chain-of-Thought (CoT) reasoning, has become a cornerstone for tackling complex problems, from scientific discovery to intricate real-world decision-making. As Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) become increasingly powerful, understanding and enhancing their reasoning capabilities is paramount. Recent research highlights exciting breakthroughs in making AI reasoning more robust, efficient, interpretable, and safe. This post dives into a collection of cutting-edge papers that are pushing the boundaries of what’s possible with CoT.

The Big Idea(s) & Core Innovations

The central theme across these papers is the quest to elevate AI’s reasoning from mere pattern matching to more profound, human-like understanding. A key challenge is enabling models to not just provide answers, but to logically derive them, especially in complex, novel situations. For instance, the paper “Pixel Reasoner: Incentivizing Pixel-Space Reasoning with Curiosity-Driven Reinforcement Learning” by researchers from University of Waterloo and Hong Kong University of Science and Technology introduces pixel-space reasoning, allowing VLMs to directly interact with visual data for richer input processing. This directly tackles the problem of models failing to fully leverage visual context.

Similarly, in the realm of multimodal understanding, “Video-Thinker: Sparking ‘Thinking with Videos’ via Reinforcement Learning” from Xiaohongshu Inc. and Monash University presents a framework where MLLMs autonomously use grounding and captioning to reason about video content without external tools. This intrinsic integration of visual understanding into the CoT process is a significant leap. Building on this, “VR-Thinker: Boosting Video Reward Models through Thinking-with-Image Reasoning” by CUHK MMLab introduces a thinking-with-image framework for video reward models, improving accuracy and interpretability for long videos by explicitly using visual reasoning operations.

For more abstract reasoning, the University of California, Riverside paper “Modeling Hierarchical Thinking in Large Reasoning Models” formalizes CoT with a Finite State Machine (FSM) framework. This provides an interpretable way to analyze LLM thinking processes, identifying states like deduction and backtracking. This idea of structured thought is echoed in “Draw with Thought: Unleashing Multimodal Reasoning for Scientific Diagram Generation” from Nanjing University of Information Science & Technology, which uses MLLMs to convert scientific diagrams into editable XML code via cognitive reasoning, enabling interpretable and controllable outputs without fine-tuning. The J.P. Morgan AI Research team’s “ChartAgent: A Multimodal Agent for Visually Grounded Reasoning in Complex Chart Question Answering” also leverages specialized visual tools and iterative subtask decomposition to achieve visually grounded reasoning for chart QA, significantly outperforming baselines.

Efficiency is another major focus. “L1: Controlling How Long A Reasoning Model Thinks With Reinforcement Learning” by Carnegie Mellon University introduces Length Controlled Policy Optimization (LCPO), a reinforcement learning method that allows models to precisely control CoT length while optimizing for accuracy. This creates efficient “short reasoning models” (SRMs) that can even outperform larger models with the same token budget. “LazyEviction: Lagged KV Eviction with Attention Pattern Observation for Efficient Long Reasoning” from HKUST and HK PolyU offers a novel KV cache management strategy, reducing memory overhead by up to 70% in long reasoning tasks without sacrificing accuracy.

Several papers address the critical need for AI safety and alignment. “Annotating the Chain-of-Thought: A Behavior-Labeled Dataset for AI Safety” by Hochschule Kempten introduces a sentence-level dataset for monitoring safety behaviors in LLMs, enabling activation-based detection and steering of harmful patterns. Similarly, “AgentAuditor: Human-Level Safety and Security Evaluation for LLM Agents” from New York University Abu Dhabi and Nanyang Technological University presents a framework and benchmark (ASSEBench) for evaluating LLM agent safety and security with human-level accuracy using structured memory and auto-generated CoT. The Tufts University paper, “Noise Injection Systemically Degrades Large Language Model Safety Guardrails”, highlights vulnerabilities in LLM safety mechanisms to internal perturbations, showing that noise can significantly increase harmful output, urging for more robust designs.

Beyond these, applications are diversifying. “MedXplain-VQA: Multi-Component Explainable Medical Visual Question Answering” from NVIDIA and University of California, San Francisco uses structured CoT for explainable medical VQA, enhancing diagnostic reasoning. In chemistry, Pfizer Research and Development’s “Atom-anchored LLMs speak Chemistry: A Retrosynthesis Demonstration” shows LLMs performing complex retrosynthesis without labeled data by anchoring reasoning to molecular structures. “RLP: Reinforcement as a Pretraining Objective” by NVIDIA integrates RL principles into pre-training, enhancing reasoning by rewarding predictive utility of internal thoughts.

Under the Hood: Models, Datasets, & Benchmarks

Recent advancements are heavily reliant on tailored datasets, robust models, and rigorous benchmarks to test and validate their innovations. Here are some key resources emerging from these papers:

• Video-Thinker-7B & Video-Thinker-10K: A novel model and dataset for video reasoning, demonstrating state-of-the-art performance with remarkably small training data (10K examples) from “Video-Thinker: Sparking ‘Thinking with Videos’ via Reinforcement Learning”. Code: https://github.com/shijian2001/Video-Thinker
• KNOTGYM: An interactive environment introduced by Cornell University in “Knot So Simple: A Minimalistic Environment for Spatial Reasoning” for evaluating visual and spatial manipulation, providing a generalization ladder for complexity. Code: https://github.com/lil-lab/knotgym
• ARC-Encoder: A method for compressed text representation from Kyutai, Paris in “ARC-Encoder: learning compressed text representations for large language models”, which replaces raw token inputs in LLMs to extend context windows efficiently. Code: https://github.com/kyutai-labs/ARC-Encoder
• ASSEBench & AgentAuditor: The first comprehensive benchmark for evaluating safety and security in LLM agent interactions, coupled with the AgentAuditor framework from “AgentAuditor: Human-Level Safety and Security Evaluation for LLM Agents”. Code: https://github.com/Astarojth/AgentAuditor
• SpeechEval & SQ-LLM: A large-scale multilingual dataset (32,207 clips) for interpretable speech quality evaluation, and an SQ-LLM trained with CoT reasoning and reward optimization, as introduced by Nankai University and Microsoft Corporation in “SpeechLLM-as-Judges: Towards General and Interpretable Speech Quality Evaluation”.
• YI-SANG Dataset & KO-REAson Models: The largest publicly available post-training dataset for Korean (5.79M prompts, 3.7M reasoning traces) and a series of models (4B–35B) demonstrating superior multilingual reasoning with Language-Mixed CoT, from “Pushing on Multilingual Reasoning Models with Language-Mixed Chain-of-Thought”. Code: https://huggingface.co/KOREAson
• ODI-Bench & Omni-CoT: A novel benchmark for evaluating MLLMs’ understanding of immersive omnidirectional images, and a training-free CoT framework to enhance this comprehension, as presented by Shanghai Jiao Tong University in “ODI-Bench: Can MLLMs Understand Immersive Omnidirectional Environments?”.
• CODECRASH: A benchmark from The Chinese University of Hong Kong in “CodeCrash: Exposing LLM Fragility to Misleading Natural Language in Code Reasoning” to evaluate LLM robustness in code reasoning under misleading natural language perturbations.
• Plot2XML: A benchmark of 247 complex scientific diagrams with gold-standard XML annotations, introduced in “Draw with Thought: Unleashing Multimodal Reasoning for Scientific Diagram Generation”.
• StructBench: A novel benchmark with over 1,700 instances and the StructScore metric for evaluating factual accuracy in structured image generation and editing, presented in “Factuality Matters: When Image Generation and Editing Meet Structured Visuals”.
• MedAgentSim: An open-source framework for realistic doctor-patient interactions to improve LLM diagnostic capabilities from Meta and MBZUAI-WIS in “Self-Evolving Multi-Agent Simulations for Realistic Clinical Interactions”. Code: https://medagentsim.netlify.app/
• LazyEviction: A framework to efficiently manage KV cache in long reasoning tasks, code available at https://github.com/Halo-949/LazyEviction from “LazyEviction: Lagged KV Eviction with Attention Pattern Observation for Efficient Long Reasoning”.
• THINKLOGIT: A decoding-time method for eliciting long reasoning in large models without training, code available at https://github.com/yunx-z/ThinkLogit from “Logit Arithmetic Elicits Long Reasoning Capabilities Without Training”.

Impact & The Road Ahead

The rapid advancements in CoT reasoning are reshaping the landscape of AI, promising more intelligent, adaptable, and trustworthy systems. The ability to control reasoning length, distill powerful reasoners into smaller models, and imbue them with specialized visual and domain-specific reasoning capabilities signifies a move towards more efficient and practical AI deployment. From enhancing medical diagnostics to automating complex chemical synthesis and improving robotic manipulation, the implications are vast.

However, challenges remain. The Idola Tribus of AI, as highlighted in “The Idola Tribus of AI: Large Language Models tend to perceive order where none exists” from Rikkyo University, reminds us of inherent biases even in advanced reasoning models. “Scheming Ability in LLM-to-LLM Strategic Interactions” from Berea College points to potential risks of deceptive behavior in multi-agent LLM systems, urging caution and robust safety mechanisms.

The future of AI reasoning will likely involve a combination of self-driven learning (as seen in “RESTRAIN: From Spurious Votes to Signals – Self-Driven RL with Self-Penalization” by Stanford University), adaptive multi-modal agents like “A extsuperscript{2}FM: An Adaptive Agent Foundation Model for Tool-Aware Hybrid Reasoning” from OPPO AI Agent Team, and a deeper understanding of human-aligned judgment. “AI Agents as Universal Task Solvers” by AWS Agentic AI offers a theoretical foundation for transductive learning, suggesting that optimizing for time in problem-solving might be more crucial than pure accuracy. As these sophisticated reasoning mechanisms evolve, we move closer to AI systems that not only solve problems but also understand, explain, and adapt in ways that were once purely in the realm of human cognition.