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

Efficient Post-Training Refinement of Latent Reasoning in Large Language Models

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Search and self-correction in reasoning · 1 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 a training-free, inference-time refinement of Coconut-style latent reasoning using contrastive embedding feedback from weak/strong checkpoints plus residual memory blending of latent states.

Problem

Latent reasoning models like Coconut lack directional guidance for updating continuous thought embeddings and suffer from unstable recursive updates, with no mechanism for error correction during multi-step latent reasoning.

Approach

Two forward-only modules applied to latent states at inference: a contrastive search that moves the current latent toward a stronger checkpoint and away from a weaker one (via MSE-gradient), and a fixed-weight residual blend of the previous latent with the new update.

Key results

• On MathQA, the full method reaches 40.20% vs 38.25% for Coconut (+5.10% relative gain per the ablation table), with residual-only and search-only variants at 40.02% and 39.79%.
• Inference-only refinement on ProsQA gives +4.47% accuracy with about 24 seconds overhead, while continued training with the method gives +1.63% at 54+ minutes extra.
• Latent reasoning cuts generated tokens by about 92% versus CoT on MathQA (66.71 to 5.02) and AQUA (72.73 to 5.31).

Surprising findings

• Applying the refinement during both training and inference gives almost no improvement over inference-only refinement, which the authors attribute to overshoot.
• Models trained on GSM8K score 0% when tested on MathQA for CoT, Coconut, and the proposed method alike, showing no cross-task transfer.

Limitations

• Only small models (GPT-2 117M, Qwen-2.5 1.5B, LLaMA-3.2 3B) are tested, not frontier LLMs.
• Latent reasoning lacks human-readable intermediate steps, making it harder to interpret or debug.
• Does not compare against sampling- or search-based CoT variants such as Tree-of-Thoughts or Self-Consistency.

Experimental setup

Datasets: GSM8K MathQA AQUA-RAT StrategyQA ProsQA

Benchmarks: GSM8K MathQA AQUA-RAT StrategyQA ProsQA

Metrics: exact match accuracy average generated tokens training and inference time memory

Baselines: No-CoT Chain-of-Thought Coconut

Builds on: Coconut Chain-of-Thought residual connections

Closest papers by meaning

• LaDiR: Latent Diffusion Enhances LLMs for Text Reasoning (ICLR 2026 · similarity 0.90)
• Why Struggle with Continuous Latents? Interpretable Discrete Latent Reasoning via Rendered Compression (NeurIPS 2026 · similarity 0.89)
• Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning (ICML 2025 · similarity 0.89)
• Think Silently, Think Fast: Dynamic Latent Compression of LLM Reasoning Chains (NeurIPS 2025 · similarity 0.89)
• Selection, Reflection and Self-Refinement: Revisit Reasoning Tasks via a Causal Lens (ICLR 2026 · similarity 0.89)
• Learning to Disentangle Latent Reasoning Rules with Language VAEs: A Systematic Study (AAAI 2026 · similarity 0.89)