Research map / Large language models / Reasoning and chain of thought

Boosting Inference with Guided Reasoning: Stochastic Exploration for Recursive Models

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes guided stochastic exploration for frozen Tiny Recursive Models: stochastic perturbations of the latent recursion form a particle cloud, reweighted online by the existing Q-head via a Feynman–Kac tilt, raising Sudoku-Extreme exact-solve from 85.9% to 98.0% without retraining.

Problem

Tiny recursive reasoning models use deterministic latent recursion at inference, leaving unexplored trajectory-level headroom; it is unclear how extra test-time compute should be allocated across alternative latent trajectories and when this will help.

Approach

Interprets recursive inference as approximate inference over latent trajectories (deterministic recursion as the one-particle, zero-noise limit); adds noise to the inner recursion as a proposal kernel and uses a sequential Monte Carlo particle filter with the existing early-stopping Q-head as a tilting potential, plus label-free diagnostics (tube stability, Q-spread bound, token entropy).

Key results

• On Sudoku-Extreme test, guided MAP improves exact-solve accuracy from 85.9% (TRM) to 98.0% without retraining.
• On Sudoku-Extreme deterministic-failure split, guided MAP reaches 85.9% vs 0.0% for TRM.
• On Maze-Hard, guided MAP stays at 85.3% (baseline 86.6%) despite oracle reaching 90.9%, as the Q-spread diagnostic predicted.

Surprising findings

• On Maze-Hard, the learned Q-head is too flat to exploit available oracle headroom, so guided selection returns to baseline and no beta or S setting helps.
• Guided MAP reaches the deterministic terminal accuracy about 3.1x earlier in training checkpoints.

Limitations

• Maze-Hard gains do not materialise; the Q-head guide is misaligned there.
• Only two benchmarks evaluated; Maze-Hard hyperparameters transferred without validation set.
• Continuum limit not pursued; rollout depth short (M<=6).

Experimental setup

Datasets: Sudoku-Extreme Maze-Hard

Benchmarks: Sudoku-Extreme Maze-Hard

Metrics: exact-solve accuracy AUROC oracle gap

Baselines: TRM deterministic unguided particle MAP oracle particle

Builds on: Tiny Recursive Model Hierarchical Reasoning Model twisted SMC Adaptive Computation Time

Closest papers by meaning

• Modeling Hierarchical Thinking in Large Reasoning Models (ICML 2026 · similarity 0.89)
• Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning (ICML 2026 · similarity 0.88)
• Selection, Reflection and Self-Refinement: Revisit Reasoning Tasks via a Causal Lens (ICLR 2026 · similarity 0.88)
• Tracing the Traces: Latent Temporal Signals for Efficient and Accurate Reasoning (ICLR 2026 · similarity 0.88)
• Efficient Reasoning via Constrained Optimization in Latent Space (NeurIPS 2026 · similarity 0.88)
• DTS: Enhancing Large Reasoning Models via Decoding Tree Sketching (ICML 2026 · similarity 0.88)