Efficient Reasoning Exploration via State-Conditioned Latent Steering with Progress Guidance
On the map: LLM reasoning → Test-time compute and efficient thinking also fits: Reinforcement learning for reasoning
Abstract (arXiv)
Best-of-N is a widely used inference strategy for complex reasoning, whose effectiveness depends on whether sampled candidates can cover diverse and high-quality reasoning paths. However, post-trained reasoning models often suffer from exploration collapse, where independent rollouts repeatedly follow similar reasoning paths and limit the gains from increasing the rollout budget. Existing methods alleviate this issue by promoting broader exploration, but do not explicitly guide exploration toward continuations that make meaningful progress, resulting in limited exploration efficiency. To address this, we propose emphunderlineState-conditioned underlineProgress-guided underlineSteering (SPS), a training-free latent steering framework. Specifically, SPS constructs a state-conditioned Direction Bank containing multiple progress-guided steering vectors for different prefix-state regions. During online inference, SPS retrieves a suitable steering vector based on the current prefix state and applies it at high-uncertainty transitions to guide the next reasoning step toward meaningful progress. Extensive experiments across multiple model scales and benchmarks demonstrate that SPS consistently outperforms strong baselines. Further analyses validate the effectiveness of its key designs and offer valuable insights for future research. The code is available at https://github.com/rattlesnakey/SPS.
Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning. Builds on 31 of its 49 references that are in the map. First posted on arXiv 21 Sep 2026.
Builds on
31 of 49 references.
Closest ancestors · 2
Same direction, newest first
- 2025Training Large Language Models To Reason In Parallel With Global Forking Tokens1 Oct
- 2025Rethinking Fine-Tuning when Scaling Test-Time Compute: Limiting Confidence Improves Mathematical Reasoning11 Feb
Other references · 29
- 2023Efficient Memory Management for Large Language Model Serving with PagedAttentioncited by 5,767
- 2023Let's Verify Step by Stepcited by 3,598
- 2023GPQA: A Graduate-Level Google-Proof Q&A Benchmarkcited by 2,748
- 2024LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Codecited by 1,906
- 2025Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Modelscited by 1,175
- 2024OlympiadBench: A Challenging Benchmark for Promoting AGI with Olympiad-Level Bilingual Multimodal Scientific Problemscited by 1,099
- 2025Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?cited by 948
- 2024Large Language Monkeys: Scaling Inference Compute with Repeated Samplingcited by 946
- 2021Did Aristotle Use a Laptop? A Question Answering Benchmark with Implicit Reasoning Strategiescited by 736
- 2025Reasoning with Exploration: An Entropy Perspectivecited by 277
- 2025MathArena: Evaluating LLMs on Uncontaminated Math Competitionscited by 275
- 2026The Assistant Axis: Situating and Stabilizing the Default Persona of Language Modelscited by 76
17 more
- 2025AceReason-Nemotron 1.1: Advancing Math and Code Reasoning through SFT and RL Synergycited by 75
- 2025LLMs Encode Harmfulness and Refusal Separatelycited by 75
- 2025SwS: Self-aware Weakness-driven Problem Synthesis in Reinforcement Learning for LLM Reasoningcited by 39
- 2025Refusal Direction is Universal Across Safety-Aligned Languagescited by 37
- 2025Unlocking General Long Chain-of-Thought Reasoning Capabilities of Large Language Models via Representation Engineeringcited by 34
- 2025TreeReview: A Dynamic Tree of Questions Framework for Deep and Efficient LLM-based Scientific Peer Reviewcited by 31
- 2025Chain-of-Reasoning: Towards Unified Mathematical Reasoning in Large Language Models via a Multi-Paradigm Perspectivecited by 29
- 2025Diversity-Incentivized Exploration for Versatile Reasoningcited by 28
- 2024Dipper: Diversity in Prompts for Producing Large Language Model Ensembles in Reasoning taskscited by 22
- 2025Improving Reasoning Performance in Large Language Models via Representation Engineeringcited by 21
- 2025Semantic-Space Exploration and Exploitation in RLVR for LLM Reasoningcited by 17
- 2025GuidedSampling: Steering LLMs Towards Diverse Candidate Solutions at Inference-Timecited by 5
- 2026Residual Decoding: Mitigating Hallucinations in Large Vision-Language Models via History-Aware Residual Guidancecited by 5
- 2026Beyond Outliers: A Data-Free Layer-wise Mixed-Precision Quantization Approach Driven by Numerical and Structural Dual-Sensitivitycited by 5
- 2026Training LLMs for Divide-and-Conquer Reasoning Elevates Test-Time Scalabilitycited by 4
- 2026Restoring Exploration after Post-Training: Latent Exploration Decoding for Large Reasoning Modelscited by 3
- 2026Massive Activations in Hybrid Linear Attention Large Language Models: Pre-Attention Spikes and Inter-Spike Plateauscited by 1
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2026Boosting Inference with Guided Reasoning: Stochastic Exploration for Recursive Models
- 2026Learning Structured Reasoning via Tractable Trajectory Control
- 2026GTS: Inference-Time Scaling of Latent Reasoning with a Learnable Gaussian Thought Sampler
- 2025Learning to Ponder: Adaptive Reasoning in Latent Space
- 2026Latent Reward Steering: An Adaptive Inference-Time Framework that Implicitly Promotes Cognitive Behaviors in Reasoning LLMs
- 2026Steering Recurrent Reasoners at Inference Time with Readout Feedback
- 2026Reliable Control-Point Selection for Steering Reasoning in Large Language Models
- 2025TrajSelector: Harnessing Latent Representations for Efficient and Effective Best-of-N in Large Reasoning Model