System 2 Reasoning for Large Language Models

This direction is split into other directions: dual-process reasoning gap. Its last version stays readable and citable here.

Version 1 of 1234 papers259 conceptsfrozen 2026-10-04

Standard Large Language Models (LLMs) are remarkably proficient at "System 1" tasks—the fast, intuitive, and associative pattern matching required for fluent conversation. However, as noted by Zhixuan Chu and Yan Wang in Professional Agents, these models often stumble when faced with "System 2" tasks: the slow, deliberate, and multi-step logical reasoning required for professional-grade problem solving.

The core of this research direction is the transition from autoregressive next-token prediction to structured, reflective deliberation. By implementing search-based architectures and process-oriented supervision, researchers are enabling models to "think" before they speak.

From Linear Chains to Search Trees

The fundamental limitation of standard LLMs is their linear, token-by-token generation. In a complex problem like a mathematical proof or a strategic game, an early mistake is often fatal because the model cannot easily backtrack.

To solve this, Shunyu Yao and colleagues introduced the Tree of Thoughts (ToT) framework. Instead of a single path, the model generates multiple potential "thoughts" (intermediate reasoning steps) and evaluates them. This allows the system to explore different branches of a solution, look ahead to see if a path is promising, and backtrack when it hits a dead end.

This shift transforms inference from a simple generation task into a search problem. Xidong Feng and Muning Wen take this further in TS-LLM by applying AlphaZero-style Monte Carlo Tree Search (MCTS). By using a learned value function to guide the search, the model can navigate deep reasoning trees—up to 64 steps—far exceeding the capabilities of standard prompting.

The Role of the Verifier

For search to be effective, the model needs a way to judge the quality of its intermediate steps. This is the role of the Verifier or Reward Model.

In the Cumulative Reasoning (CR) framework, Jingqin Yang and Andrew C Yao decompose the reasoning process into specialized roles:

  1. Proposers suggest new logical steps based on current premises.
  2. Verifiers check these steps for correctness.
  3. Reporters decide when the goal has been reached.

By building a Directed Acyclic Graph (DAG) of verified propositions rather than a fragile linear chain, the system becomes significantly more robust. Similarly, the STILL-1 framework by Jinhao Jiang uses a generative reward model to guide MCTS, achieving massive gains on difficult math benchmarks by explicitly rewarding correct intermediate logic rather than just the final answer.

Internalizing Deliberation

While inference-time search is powerful, it is computationally expensive. A secondary goal in this field is "internalization"—training the model to perform this deliberate reasoning within its own weights, effectively turning System 2 expertise into System 1 intuition.

The CogniDual Framework, proposed by Yongxin Deng, mimics the human learning process where slow, sequential processing eventually becomes fast and automatic. By fine-tuning models on their own successful search-based reasoning chains, researchers can "distill" the benefits of slow thinking into a model that can provide correct direct answers more efficiently.

Evaluating the Reasoning Process

As models move toward these "Large Reasoning Model" (LRM) architectures, traditional benchmarks that only check the final answer (like multiple-choice questions) become insufficient.

New benchmarks like MR-Ben focus on meta-reasoning: the ability of a model to identify, locate, and correct errors within a complex reasoning chain. This reveals a "reasoning gap"—many models can generate a correct-looking answer through pattern matching but fail to critically evaluate the logical validity of the steps taken to get there.

Why It Works: Inference-Time Scaling

The success of this direction suggests a new scaling law for AI. While previous eras focused on scaling training data and parameter counts, this research shows that scaling compute at inference time—by allowing the model more time and "thoughts" to explore a problem—can yield dramatic improvements in performance.

Whether through Collective Monte Carlo Tree Search (Mulberry) or Active Deduction (MARS-Bench), the goal remains the same: moving beyond the "first thing that comes to mind" toward a model that can simulate, verify, and plan its way through the world's most complex problems.

Papers in this version

Showing 60 of 234 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_system_2_reasoning_for_large_language_mo_2a4b8120,
  title = {System 2 Reasoning for Large Language Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/system-2-reasoning-for-large-language-models/rdr_2a4b8120e95b48a88a68dd5706cf4f51},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_system_2_reasoning_for_large_language_mo_2a4b8120_v1,
  title = {System 2 Reasoning for Large Language Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/system-2-reasoning-for-large-language-models/rdr_2a4b8120e95b48a88a68dd5706cf4f51/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}