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

Eliciting Chain-of-Thought in Base LLMs via Gradient-Based Representation Optimization

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Chain-of-thought training and analysis · 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 gradient-based hidden-state optimization method that steers base LLMs toward chain-of-thought reasoning by maximizing a classifier likelihood minus an L2 prior penalty at inference time.

Problem

Base (pre-trained only) LLMs often answer directly without chain-of-thought; existing linear activation steering uses fixed, unconstrained control vectors that cause distribution shift and degrade text quality.

Approach

Frame CoT elicitation as MAP estimation over hidden states: maximize log-likelihood from a trained MLP CoT classifier minus a lambda-weighted L2 distance to the original hidden state, with adaptive step size and Gaussian noise, applied during prefill and decoding.

Key results

• On Mistral-7b GSM8K, the method reaches 18.24% accuracy vs 11.32% vanilla and 16.83% for the best baseline (DA).
• On Gemma-7b MultiArith, accuracy improves from 30.55% (vanilla) to 45.11%, the largest reported gain.
• Gains on commonsense and logic tasks are smaller, averaging about 1.8% over baselines per the authors.

Surprising findings

• Classifiers trained on Mistral-7B hidden states transfer to other models (Gemma, Qwen, Phi) without retraining per model, according to the authors.
• Increasing the control-vector strength for C-LR eventually drops performance below greedy decoding and degrades language quality.

Limitations

• Can only unlock reasoning latent in the base model; cannot inject new reasoning capability.
• Effectiveness depends on classifier quality and layer/hidden-state selection.
• Commonsense and logic gains are modest; the authors attribute this to needed domain knowledge.

Experimental setup

Datasets: GSM8K MultiArith SVAMP MAWPS CommonsenseQA StrategyQA Objects Tracking Coin Flip

Benchmarks: GSM8K MultiArith SVAMP MAWPS CommonsenseQA StrategyQA Objects Tracking Coin Flip

Metrics: answer accuracy fluency (n-gram entropy) perplexity GPT-4o judge reasoning score

Baselines: C-DiM C-PCA C-LR P-SVM Directional Ablation

Builds on: representation engineering contrastive activation addition directional ablation top-k start greedy decoding

Closest papers by meaning

• How Far Ahead Do LLMs Plan? Uncovering the Latent Horizon in Chain-of-Thought Reasoning (ICML 2026 · similarity 0.92)
• Markovian Transformers for Informative Language Modeling (ICLR 2026 · similarity 0.92)
• C3oT: Generating Shorter Chain-of-Thought Without Compromising Effectiveness (AAAI 2025 · similarity 0.91)
• SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought (NeurIPS 2025 · similarity 0.91)
• System-1.5 Reasoning: Traversal in Language and Latent Spaces with Dynamic Shortcuts (NeurIPS 2025 · similarity 0.91)
• Deep Hidden Cognition Facilitates Reliable Chain-of-Thought Reasoning (AAAI 2026 · similarity 0.91)