Soft Reasoning: Navigating Solution Spaces in Large Language Models through Controlled Embedding Exploration
On the map: LLM reasoning → Test-time compute and efficient thinking
Abstract (arXiv)
Large Language Models (LLMs) struggle with complex reasoning due to limited diversity and inefficient search. We propose Soft Reasoning, an embedding-based search framework that optimises the embedding of the first token to guide generation. It combines (1) embedding perturbation for controlled exploration and (2) Bayesian optimisation to refine embeddings via a verifier-guided objective, balancing exploration and exploitation. This approach improves reasoning accuracy and coherence while avoiding reliance on heuristic search. Experiments demonstrate superior correctness with minimal computation, making it a scalable, model-agnostic solution. The code is released at https://github.com/alickzhu/Soft-Reasoning.
Key facts. Cited by 7 papers in the map (data through 14 Oct 2026). Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning. Builds on 7 of its 52 references that are in the map. First posted on arXiv 30 May 2025.
Builds on
7 of 52 references.
References · 7
- 2021Training Verifiers to Solve Math Word Problemscited by 10,109
- 2023Mistral 7Bcited by 3,164
- 2024Qwen2 Technical Reportcited by 2,120
- 2024Large Language Monkeys: Scaling Inference Compute with Repeated Samplingcited by 991
- 2018A Tutorial on Bayesian Optimizationcited by 744
- 2019Neural Text Generation with Unlikelihood Trainingcited by 503
- 2025Towards Large Reasoning Models: A Survey of Reinforced Reasoning with Large Language Modelscited by 293
Built on by
Citing papers per year: 2025: 5; 2026: 2.
7 papers.
- 2025A Survey on Latent Reasoningcited by 81
- 2025Online Multi-LLM Selection via Contextual Bandits under Unstructured Context Evolutioncited by 19
- 2025Semantic-guided Diverse Decoding for Large Language Modelcited by 9
- 2025Beyond Perplexity: Let the Reader Select Retrieval Summaries via Spectrum Projection Scorecited by 1
- 2025Test-time reward-guided alignment of language models by importance sampling on pre-logit spacecited by 0
- 2026Partial Reasoning in Language Models: Search and Refinement Guided by Uncertaintycited by 0
- 2026Beyond Policy Rollouts: Embedding-Space Directional Exploration with Anchored Noise for LLM RLVRcited by 0
- 2026Beyond Policy Rollouts: Embedding-Space Directional Exploration with Anchored Noise for LLM RLVR12 Oct 2026
- 2026Partial Reasoning in Language Models: Search and Refinement Guided by Uncertainty17 Jan 2026
- 2025Test-time reward-guided alignment of language models by importance sampling on pre-logit space30 Oct 2025
- 2025Beyond Perplexity: Let the Reader Select Retrieval Summaries via Spectrum Projection Score8 Aug 2025
- 2025A Survey on Latent Reasoning8 Jul 2025
- 2025Semantic-guided Diverse Decoding for Large Language Model30 Jun 2025
- 2025Online Multi-LLM Selection via Contextual Bandits under Unstructured Context Evolution21 Jun 2025
Similar, not cited
Close in wording, no citation either way
- 2026SeLaR: Selective Latent Reasoning in Large Language Models
- 2026Reasoning Quality Matters: Combating Reasoning Collapse in LLM-based Embedding Learning
- 2025Improving Latent Reasoning in LLMs via Soft Concept Mixing
- 2026A Model with No Head and Many Thoughts
- 2025Search-R3: Unifying Reasoning and Embedding in Large Language Models
- 2025Advancing Reasoning in Large Language Models: Promising Methods and Approaches
- 2026Test-Time Compute for Frozen Embedding Models through Agentic Program Search
- 2026Exploration-Driven Optimization for Test-Time Large Language Model Reasoning