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

Lookahead Path Likelihood Optimization for Diffusion LLMs

ICML 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 Path LL as a proxy for dLLM generation quality, an optimistic lookahead value estimator (POKE), and POKE-SMC, an SMC search over unmasking orders that improves LLaDA accuracy by about 2-3 points at matched budget.

Problem

Inference performance of diffusion LLMs depends strongly on the unmasking order, but heuristics greedily optimize local confidence and provide no global guidance on which decoding paths are accurate.

Approach

Defines path log-likelihood (the joint likelihood of a sequence under its unmasking trajectory) as the optimization target, estimates expected future Path LL with POKE (random rollouts with product likelihood plus an entropy-based total-correlation correction), and uses it as the SMC importance weight for periodic resampling.

Key results

• POKE-SMC (M=4) reaches 47.7 average vs 44.8 for PC-Sampler on LLaDA-8B-Instruct (+2.9), and 49.0 on LLaDA-1.5-8B (+2.0).
• Under 2 tokens per step, POKE-SMC gains 5.4 points average over PC-Sampler on LLaDA-8B-Instruct.
• POKE-SMC beats E-SMC (45.6 avg) and Majority Voting at comparable latency (~42s per sample) on LLaDA-8B-Instruct.

Surprising findings

• Entropy-based heuristics such as Entropy and Confidence collapse on Sudoku (e.g., Semi-AR + Conf 0.0), while Margin excels there, showing heuristic inconsistency across tasks.
• ReMDM frequently fails to beat single-particle baselines despite about 2x compute.

Limitations

• Evaluation limited to 8B-scale LLaDA models; larger dLLMs such as LLaDA2.0 left to future work.
• Roughly 4x inference overhead from maintaining M=4 particles relative to the single-particle baseline.
• Resampling strategy is simple periodic resampling at fixed intervals.

Experimental setup

Datasets: GSM8K MATH500 HumanEval MBPP Sudoku Countdown

Benchmarks: HumanEval MBPP GSM8K MATH500 Countdown Sudoku

Metrics: accuracy latency (s per sample) average accuracy

Baselines: Uniform Confidence Entropy Margin EB-Sampler Semi-AR PC-Sampler Majority Voting ReMDM E-SMC LookUM

Builds on: E-SMC PC-Sampler Sequential Monte Carlo Monte Carlo tree search

Closest papers by meaning

• Efficient Test-Time Scaling via Hierarchical Search and Self-Verification for Discrete Diffusion Language Models (ICML 2026 · similarity 0.90)
• Fast-dLLM++: Fr\'{e}chet Profile Decoding for Faster Diffusion LLM Inference (NeurIPS 2026 · similarity 0.90)
• Self-Rewarding Sequential Monte Carlo for Masked Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
• UnMaskFork: Test-Time Scaling for Masked Diffusion via Deterministic Action Branching (ICML 2026 · similarity 0.89)
• d2: Improving Reasoning in Diffusion Language Models via Trajectory Likelihood Estimation (ICML 2026 · similarity 0.89)
• Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs (NeurIPS 2026 · similarity 0.89)