paper-with-me

Papers

Therefore I am. I Think

2026-04-01 · Esakkivel Esakkiraja, Sai Rajeswar, Denis Akhiyarov, Rajagopal Venkatesaramani arxiv

We consider the question: when a large language reasoning model makes a choice, did it think first and then decide to, or decide first and then think? In this paper, we present evidence that detectable, early-encoded decisions shape chain-of-thought in reasoning models. Specifically, we show that a simple linear probe successfully decodes tool-calling decisions from pre-generation activations with very high confidence, and in some cases, even before a single reasoning token is produced. Activation steering supports this causally: perturbing the decision direction leads to inflated deliberation, and flips behavior in many examples (between 7 - 79% depending on model and benchmark). We also show through behavioral analysis that, when steering changes the decision, the chain-of-thought process often rationalizes the flip rather than resisting it. Together, these results suggest that reasoning models can encode action choices before they begin to deliberate in text.

📄 PDF Abstract BibTeX arXiv:2604.01202

Code (0)

등록된 구현이 없습니다.

Similar Papers 제목 키워드 기반

Fast-Slow-Thinking: Complex Task Solving with Large Language Models

2025-04-11 · Yiliu Sun, Yanfang Zhang, Zicheng Zhao, Sheng Wan 외

Nowadays, Large Language Models (LLMs) have been gradually employed to solve complex tasks. To face the challenge, task decomposition has become an effective way, which proposes to divide a complex task into multiple sim…

Distilling Reasoning Ability from Large Language Models with Adaptive Thinking

2024-04-14 · Xiaoshu Chen, Sihang Zhou, Ke Liang, Xinwang Liu

Chain of thought finetuning (cot-finetuning) aims to endow small language models (SLM) with reasoning ability to improve their performance towards specific tasks by allowing them to imitate the reasoning procedure of lar…

Hallucination

ThinkPrune: Pruning Long Chain-of-Thought of LLMs via Reinforcement Learning

2025-04-02 · Bairu Hou, Yang Zhang, Jiabao Ji, Yujian Liu 외

We present ThinkPrune, a simple yet effective method for pruning the thinking length for long-thinking LLMs, which has been found to often produce inefficient and redundant thinking processes. Existing preliminary explor…

Reinforcement Learning (RL)

How and What to Imagine? Visual Thinking in Unified Multimodal Models for Cross-View Spatial Reasoning

2026-05-26 · Qian Yang, Ankur Sikarwar, Huy Le, Le Zhang 외 arxiv

Cross-view spatial reasoning remains a weak spot for vision-language models (VLMs): they often reason in language and lose the fine-grained geometry needed for the task. Thinking with images aims to address this by gener…

Domain GeneralizationSpatial ReasoningText Generation

Enhance Reasoning Ability of Visual-Language Models via Large Language Models

2023-05-22 · Yueting Yang, Xintong Zhang, Wenjuan Han

Pre-trained visual language models (VLM) have shown excellent performance in image caption tasks. However, it sometimes shows insufficient reasoning ability. In contrast, large language models (LLMs) emerge with powerful…

Language ModelingLanguage ModellingLarge Language Model