# Multi-step reasoning

Test-time search for multi-step LLM reasoning has long traded between two imperfect signals: post-hoc voting keeps diversity but cannot steer generation toward correct paths, while beam-style verifier guidance intervenes early but over-concentrates and amplifies verifier errors. VGSC multiplies step-level correctness with a future-outcome prediction into an acceptance weight, applies bounded rejection-resampling to candidate steps during decoding, and keeps multiple trajectories for final self-consistency voting. Across five reasoning benchmarks with the Qwen2.5 Math 1.5B Instruct generator and Qwen2.5 Math PRM 7B verifier, VGSC reaches 63.1% average accuracy at verifier-trust exponent 10, above the strongest non-VGSC baseline BS+SC-Future at 60.7%.

Test-time search for multi-step LLM reasoning has long traded between two imperfect signals: post-hoc voting keeps diversity but cannot steer generation toward correct paths, while beam-style verifier guidance intervenes early but over-concentrates and amplifies verifier errors. VGSC multiplies step-level correctness with a future-outcome prediction into an acceptance weight, applies bounded rejection-resampling to candidate steps during decoding, and keeps multiple trajectories for final self-consistency voting. Across five reasoning benchmarks with the Qwen2.5 Math 1.5B Instruct generator and Qwen2.5 Math PRM 7B verifier, VGSC reaches 63.1% average accuracy at verifier-trust exponent 10, above the strongest non-VGSC baseline BS+SC-Future at 60.7%.

Molecular cloning demands end-to-end reasoning, from picking starting materials and designing sequences to planning a series of reactions, while models have been reliable only on individual steps. BenchBench-Cloning tests that end-to-end ability with 65 cloning tasks drawn from real experiments, split into protocol development and construct design. GPT-6 Astra solved 30.6% of tasks on a single attempt, ahead of GPT-6.1 Sol at 22.2% and Claude Opus 5.5 at 20.5%; the same top model solved 72% of protocol-development tasks within ten attempts, against 27% on one.

Molecular cloning demands end-to-end reasoning, from picking starting materials and designing sequences to planning a series of reactions, while models have been reliable only on individual steps. BenchBench-Cloning tests that end-to-end ability with 65 cloning tasks drawn from real experiments, split into protocol development and construct design. GPT-6 Astra solved 30.6% of tasks on a single attempt, ahead of GPT-6.1 Sol at 22.2% and Claude Opus 5.5 at 20.5%; the same top model solved 72% of protocol-development tasks within ten attempts, against 27% on one.

HyperThink proposes a text-to-parameter approach in which a lightweight hypernetwork reads the question and predicts updates to a small subset of the base LLM's parameters, while a vector-quantized decoder constrains those updates to a finite set of reusable patterns; trained end-to-end on outputs from the base model itself, it eliminates long thinking traces at test time, so that after one hypernetwork forward pass the adapted model generates a concise step-by-step solution and final answer without an intermediate trace, using far fewer tokens while retaining strong reasoning performance, and it improves the low-latency region of the accuracy-latency trade-off on mathematical and general reasoning tasks, with its strongest gains in the near-non-thinking regime.

HyperThink proposes a text-to-parameter approach in which a lightweight hypernetwork reads the question and predicts updates to a small subset of the base LLM's parameters, while a vector-quantized decoder constrains those updates to a finite set of reusable patterns; trained end-to-end on outputs from the base model itself, it eliminates long thinking traces at test time, so that after one hypernetwork forward pass the adapted model generates a concise step-by-step solution and final answer without an intermediate trace, using far fewer tokens while retaining strong reasoning performance, and it improves the low-latency region of the accuracy-latency trade-off on mathematical and general reasoning tasks, with its strongest gains in the near-non-thinking regime.