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

An Architecture Search Framework for Inference-Time Techniques

ICML 2025 · 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.

Introduces Archon, a modular framework that uses Bayesian optimization over a space of LLM ensembles, sampling, fusers, critics, rankers, and verifiers to automatically design inference-time architectures that beat frontier models and prior inference-time systems on instruction-following, reasoning, and coding benchmarks.

Problem

Best practices for combining inference-time techniques (repeated sampling, ensembling, fusion, ranking, critiquing, verification, unit tests) across models and tasks are underdeveloped, and the design space of such inference-time architectures is too large to explore manually.

Approach

Treat inference-time architecture as a layered, neural-network-like pipeline of LLM components (generator, fuser, critic, ranker, verifier, unit-test generator/evaluator) with no learned weights, and search its discrete configuration space with Bayesian optimization under an inference-call or token budget to produce task-specific or general-purpose architectures.

Key results

• Archon architectures outperform frontier models (o1, GPT-4o, Claude 3.5 Sonnet) and prior inference-time systems by an average of 15.1% across instruction-following, reasoning, and coding benchmarks.
• Open-source-only Archon architectures beat state-of-the-art open-source LLMs and inference systems by 11.2% on average.
• Archon uses 20.0% fewer inference calls, 15.1% fewer input tokens, and 13.5% fewer output tokens than alternative inference-time architectures; on CodeContests, GPT-4o Pass@1 rises from 40 to 58 of 140 questions with unit-test generation/evaluation and sampling.

Measured effects reported in the paper

Surprising findings

• Ensembles of generators with ranking alone often underperform the best single sampled model, e.g. Ranker-only ensembles lag the 10-sample best model by about 3 points on average.
• Verifier and unit-test components help reasoning and coding tasks but can reduce instruction-following performance; 7B-model ensembles are less effective for critique and fusion.

Limitations

• Architectures make multiple sequential calls, taking up to about 5x more time and money than a single LLM call.
• Bayesian optimization is less effective under very small inference-call budgets (<20 calls), where greedy search may match it.
• Search space is restricted to six hyperparameter axes with fixed prompts and temperature 0.7, so the design space is not fully explored.

Experimental setup

Datasets: MT-Bench AlpacaEval 2.0 Arena-Hard-Auto MixEval MixEval-Hard MATH CodeContests GPQA MMLU MMLU-Pro

Benchmarks: MT-Bench AlpacaEval 2.0 Arena-Hard-Auto MixEval MixEval-Hard MATH CodeContests GPQA Diamond MMLU MMLU-Pro

Metrics: win rate length-controlled win rate accuracy pass@1 inference calls input/output tokens

Baselines: GPT-4o Claude 3.5 Sonnet Llama 3.1 405B Instruct OpenAI o1 MoA ADAS AFlow

Builds on: Mixture-of-Agents LLM-Blender ADAS AFlow DSPy

Closest papers by meaning

• Scaling Laws Meet Model Architecture: Toward Inference-Efficient LLMs (ICLR 2026 · similarity 0.90)
• Puzzle: Distillation-Based NAS for Inference-Optimized LLMs (ICML 2025 · similarity 0.89)
• Every Rollout Counts: Optimal Resource Allocation for Efficient Test-Time Scaling (NeurIPS 2025 · similarity 0.89)
• TimeBill: Time-Budgeted Inference for Large Language Models (AAAI 2026 · similarity 0.89)
• Inference Scaling Laws: An Empirical Analysis of Compute-Optimal Inference for LLM Problem-Solving (ICLR 2025 · similarity 0.89)
• AgentTTS: Large Language Model Agent for Test-time Compute-optimal Scaling Strategy in Complex Tasks (NeurIPS 2025 · similarity 0.89)