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

GuidedSampling: Steering LLMs Towards Diverse Candidate Solutions at Inference-Time

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling · 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 GuidedSampling, which first explores diverse concepts (theorem names) for a question and then generates solutions conditioned on each concept, improving diversity and pass@k at inference and when used to create fine-tuning data.

Problem

Repeated sampling at inference time tends to produce redundant solutions that rely on the same underlying concept, limiting pass@k gains.

Approach

Decouple exploration (iteratively generate diverse concepts, each conditioned on the question and previous concepts) from generation (sample solutions conditioned on each concept), and use the resulting trajectories as synthetic post-training data (FA and CAA settings), with a theoretical condition for when it beats repeated sampling.

Key results

• GuidedSampling improves base-model pass@50 by about 16.01% on average over repeated sampling (21.8% on MATH, 11.87% on GPQA-Diamond, 11.28% on HumanEval, 3.08% on OlympiadBench).
• Llama-3.2-3B-Instruct fine-tuned with CAA trajectories reaches 60.06% pass@5 on MATH versus 44.78% for RS-trained model.
• Average concepts per instance rise from 1.67 (RS-trained) to 3.03 (CAA-trained).

Surprising findings

• GuidedSampling underperforms repeated sampling on CommonSenseQA (98.94% RS vs 95.66% GS pass@50) and on Qwen2.5-3B HumanEval, where the model generates few distinct concepts.
• Llama-3.2-3B-Instruct produces about 3.7x more unique concepts than Qwen2.5-3B-Instruct on average.

Limitations

• Evaluated only on small open models (3B) plus limited checks on GPT-4o-mini and Phi-4-mini; no proprietary frontier models.
• Performance depends on the model's ability to generate relevant, diverse concepts; fails in domains like commonsense reasoning.
• Majority voting and verifier selection do not always benefit from the added diversity.

Experimental setup

Datasets: OpenMathInstruct-2 MATH GPQA-Diamond HumanEval OlympiadBench CommonSenseQA

Benchmarks: MATH GPQA-Diamond HumanEval OlympiadBench CommonSenseQA

Metrics: pass@k pass@1 pass@5 pass@50 majority voting accuracy distinct concept count

Baselines: Repeated Sampling Tree-of-Thought STaR RandIdeaInjection BoN

Builds on: Repeated Sampling Tree-of-Thoughts STaR Self-Consistency

Closest papers by meaning

• Do We Truly Need So Many Samples? Multi-LLM Repeated Sampling Efficiently Scales Test-Time Compute (AAAI 2026 · similarity 0.89)
• DISC: Dynamic Decomposition Improves LLM Inference Scaling (NeurIPS 2025 · similarity 0.87)
• An Architecture Search Framework for Inference-Time Techniques (ICML 2025 · similarity 0.86)
• The Silent Hyperparameter: Quantifying the Impact of Inference Backends on LLM Reproducibility (NeurIPS 2026 · similarity 0.86)
• Wider or Deeper? Scaling LLM Inference-Time Compute with Adaptive Branching Tree Search (NeurIPS 2025 · similarity 0.86)
• Reasoning with Sampling: Your Base Model is Smarter Than You Think (ICLR 2026 · similarity 0.86)