System 2 Deliberative Reasoning in Large Language Models

85 concepts72 papers2 proposals

Large Language Models (LLMs) are remarkably proficient at tasks that mirror human "System 1" thinking—fast, instinctive, and associative pattern matching. However, they notoriously struggle with "System 2" tasks, which require slow, deliberate, and multi-step logical planning. This "reasoning gap" is the primary barrier preventing models from mastering complex domains like molecular structure elucidation, advanced mathematics, and strategic board games.

The Architectural Bias Toward Intuition

The fundamental limitation stems from the transformer's autoregressive nature. As Cheng Yang et al. and Rui Mao et al. observe, standard LLMs function through rapid next-token prediction. This feedforward process lacks internal mechanisms for backtracking, long-term strategic planning, or evaluating the logical consequences of a token before it is generated.

In psychological terms, Gili Lior notes that LLMs operate in a state of perpetual intuitive bias, lacking the corrective logical overrides that humans use to solve unfamiliar problems. This results in models that can write creative prose but fail at simple Sudoku or Minesweeper puzzles, as Benjamin Estermann et al. demonstrate, because these tasks require iterative reasoning rather than statistical correlation.

Bridging the Gap with Search and Planning

To move beyond instinctive responses, researchers are integrating search-based algorithms that force the model to explore multiple reasoning paths. John Schultz et al. propose two paradigms for this: External Search, where the LLM guides a Monte Carlo Tree Search (MCTS), and Internal Search, where the model is trained to perform a minimax search within its own context window.

Similarly, the RAG-Star framework by Jinhao Jiang et al. combines MCTS with retrieval-augmented generation. By allowing the model to plan reasoning steps internally and then verify them against external facts, the system reduces hallucinations and maintains logical consistency across multi-hop questions. For specialized tasks like code generation, Kounianhua Du et al. use a multi-agent tree search to explore complex reasoning paths, bridging the gap between fast execution and structured planning.

Verification as a Deliberative Layer

A key mechanism for System 2 reasoning is the ability to verify intermediate steps. Cheng Yang et al. introduced LLM2, which adds a process-based verifier to guide token generation in real-time. By training on "what-if" scenarios, the model learns to anticipate the logical outcomes of its choices.

This approach is scaled in rStar-Math by Xinyu Guan et al., where small language models (SLMs) use MCTS and code-verified reasoning steps to outperform much larger models. The intuition here is that "deep thinking" does not necessarily require more parameters, but rather a more rigorous verification of the reasoning trajectory.

Internalizing Deliberation

While search and verification are effective, they are computationally expensive. The CogniDual Framework by Yongxin Deng et al. explores how models can "internalize" these slow processes. By training on their own successful Chain-of-Thought (CoT) outputs, LLMs can eventually achieve high accuracy on logical tasks without the latency of step-by-step deliberation—mimicking the human transition from a novice's slow thought to an expert's fast intuition.

Other approaches seek to evolve these capabilities directly. Zeyuan Ma et al. propose Evolutionary Reasoning Optimization (ERO), applying "survival of the fittest" to model parameters to unlock deep logical reasoning. This allows even smaller models to tackle complex visual-logical puzzles like the Abstraction and Reasoning Corpus (ARC).

The Path to Neural-Symbolic Integration

Finally, some researchers argue that pure neural architectures may always struggle with the "symbol grounding" required for perfect logic. Yuan Lin et al. developed the Neural-Symbolic Processor (NSP), which pairs neural pattern matching with a symbolic executor to handle math and logic. Luís C. Lamb et al. suggest that Graph Neural Networks (GNNs) are particularly well-suited for this, as their relational structure naturally supports the explainable, symbolic reasoning characteristic of System 2.

By incorporating these higher-level cognitive inductive biases—such as causal reasoning and modularity—Anirudh Goyal and Yoshua Bengio argue we can move toward AI that doesn't just predict the next token, but understands the underlying rules of the world it is describing.

Key Foundational Work

The Consciousness Prior (2017) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_dual_process_reasoning_gap_48e348df,
  title = {dual-process reasoning gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dual-process-reasoning-gap/rdr_48e348df221744519e78e1e7fe9e0fd0},
  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_dual_process_reasoning_gap_48e348df_v1,
  title = {dual-process reasoning gap},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dual-process-reasoning-gap/rdr_48e348df221744519e78e1e7fe9e0fd0/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake