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

LatentChem: From Textual CoT to Latent Thinking in Chemical Reasoning

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 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 LatentChem, a chemical LLM that replaces explicit textual CoT with continuous latent thought vectors and a ChemUpdater that re-queries molecular features, trained via a four-stage protocol ending in GRPO; reports spontaneous internalization of reasoning and large token/latency savings.

Problem

Explicit chain-of-thought in chemical LLMs forces nonverbal chemical reasoning into discrete natural language ("modality mismatch"), creating a reasoning bottleneck and high inference cost.

Approach

A latent reasoning interface with continuous thought vectors fed back into the LLM, plus a ChemUpdater cross-attention that refreshes molecular tokens at each latent step; under outcome-only GRPO rewards the model drops textual CoT in favor of latent computation.

Key results

• LatentChem achieves a 59.88% non-tie win rate vs the explicit CoT baseline (Stage 1+2+4) on ChemCoTBench, and 85.26% on ChEBI-20.
• Average reasoning step overhead reduced 10.84x (5.96x measured wall-clock speedup) across benchmarks.
• On GSK3-beta molecule optimization, LatentChem reaches 82% success rate vs 67% for explicit CoT.

Surprising findings

• Under outcome-only GRPO rewards, the model spontaneously drops textual CoT and reasons in latent space in most tasks.
• When latent budget is restricted, the model reactivates explicit CoT, showing a compensatory trade-off between latent and textual reasoning.

Limitations

• Reduced interpretability of latent reasoning compared to explicit CoT.
• Gains are task-dependent with some regressions; closed-ended reaction prediction margins are narrower.
• Source-level overlap via shared public databases and sensitivity to reward design and GRPO stability.

Experimental setup

Datasets: ChemCoTDataset ChemCoTBench Mol-Instructions ChEBI-20 ChemLLMBench

Benchmarks: ChemCoTBench Mol-Instructions ChEBI-20 ChemLLMBench

Metrics: non-tie win rate success rate property improvement METEOR reasoning step overhead wall-clock speedup

Baselines: Qwen-3-8B Qwen-3-8B (SFT) Stage 1 explicit CoT Stage 1+2 explicit CoT Stage 1+2+4 explicit CoT Coconut-Chem

Builds on: Coconut GRPO Perceiver Resampler Chain-of-Thought

Closest papers by meaning

• Training a Scientific Reasoning Model for Chemistry (NeurIPS 2025 · similarity 0.92)
• Beyond Chemical QA: Evaluating LLM's Chemical Reasoning with Modular Chemical Operations (NeurIPS 2025 · similarity 0.91)
• Structured Chemistry Reasoning with Large Language Models (ICML 2024 · similarity 0.89)
• Why Struggle with Continuous Latents? Interpretable Discrete Latent Reasoning via Rendered Compression (NeurIPS 2026 · similarity 0.88)
• A Formal Comparison Between Chain of Thought and Latent Thought (ICML 2026 · similarity 0.87)
• From What to Why: A Multi-Agent System for Evidence-based Chemical Reaction Condition Reasoning (ICLR 2026 · similarity 0.87)