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

TRIM: Hybrid Inference via Targeted Stepwise Routing in Multi-Step Reasoning Tasks

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 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 TRIM, a step-level hybrid inference method that lets a cheap model generate reasoning steps and routes only PRM-flagged critical steps to a strong model, using threshold, RL (sequence and aggregated features), and POMDP routers.

Problem

Query-level LLM routing assigns an entire generation to one model, wasting expensive-model tokens on routine steps, while multi-step math reasoning suffers cascading failures from single erroneous steps.

Approach

Routes at the granularity of individual reasoning steps: a cheap model proposes each step, a process reward model scores it, and a router either accepts it or regenerates that single step with the expensive model before returning control to the cheap model. Routers range from a myopic threshold to RL policies over stepwise features and a POMDP that treats PRM scores as noisy observations of latent

Key results

• On MATH-500, TRIM-Thr achieves Delta_IBC 4.75 vs 0.95 for AutoMix-PRM (about 5x better cost efficiency); TRIM-POMDP reaches CPT(95%) at 17.98% expensive tokens.
• On AIME, TRIM-POMDP achieves Delta_IBC 5.00 versus 0.79 for SW Ranking (the best query-level baseline in that table), reported as up to 6x cost efficiency.
• TRIM-Agg trained only on AIME reaches Delta_IBC 2.57 on OlympiadBench and 3.12 on Minerva Math, while BERT and SW Ranking degrade to near zero or negative on those benchmarks.

Surprising findings

• Allowing the strong model to take over all remaining steps after one intervention (One-Step Thr) is substantially less cost-efficient than returning control to the weak model after a single step.
• Routing with the weak model reaches the strong model's performance on MATH-500 with about 80% fewer expensive-model tokens.

Limitations

• Relies on a process reward model whose accuracy is domain-specific; PRM-based routing is validated mainly on math, and generalization beyond math uses self-scoring with limited evidence (GPQA only with Qwen3 pair).
• Cost metric counts only expensive-model decode tokens and ignores prefill and cheap-model cost; a 20% prefill overhead adjustment is only discussed analytically.
• AutoMix-PRM is an authors' modification of the baseline, and the Smoothie adaptation requires a third model and a non-native cost-performance routing.

Experimental setup

Datasets: AIME MATH ProcessBench Omni-MATH PRM800K

Benchmarks: MATH-500 AIME OlympiadBench Minerva Math GPQA-Diamond

Metrics: cost-performance threshold (CPT) incremental benefit per cost (IBC) performance gap recovered (PGR) budgeted accuracy

Baselines: BERT router matrix factorization router SW ranking router Smoothie AutoMix AutoMix-PRM RouteLLM

Builds on: RouteLLM AutoMix Smoothie Reward-Guided Speculative Decoding SpecReason

Closest papers by meaning

• Confidence-Guided Stepwise Model Routing for Cost-Efficient Reasoning (AAAI 2026 · similarity 0.90)
• TrimR: Verifier-based Training-Free Thinking Trimming for Efficient Test-Time Scaling (ICLR 2026 · similarity 0.89)
• Reasoning Is Not Free: Robust Adaptive Cost-Efficient Routing for LLM-as-a-Judge (ICML 2026 · similarity 0.89)
• GLoRe: When, Where, and How to Improve LLM Reasoning via Global and Local Refinements (ICML 2024 · similarity 0.88)
• Making Slow Thinking Faster: Compressing LLM Chain-of-Thought via Step Entropy (ICLR 2026 · similarity 0.88)
• Think Smarter not Harder: Adaptive Reasoning with Inference Aware Optimization (ICML 2025 · similarity 0.88)