AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Constrained Decoding of Diffusion LLMs with Context-Free Grammars

Niels Mündler, Jasper Dekoninck, Martin Vechev

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Parallel and early-exit decoding

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add CFG constrained decoding (without completion)infilling LLMs · HumanEval C++ 1-MRI, 2-MRI, 3-MRIsyntactic correctness: +5.2%, +22.5%, +31.5% absoluteImproved
complete aborted instances by sampling from the intersection language (Con. vs Con.-) (ablation)constrained decoding · HumanEval C++ MRIsyntactic correctness: 95.8% syntactically validImproved
add constrained decodinginfilling LLMs · HumanEval C++ MRIfunctional correctness (pass@1): +2.4% (Con.-), +2.8% (Con.)Improved
add constrained decoding (without completion)diffusion LLMs · C++, JSON, SMILESsyntactic correctness: +16.1%, +14.7%, +26.0%Improved
complete aborted instances from the intersection language (Con. vs Con.-) (ablation)DreamCoder 7B · C++syntactic correctness: 19.0 -> 99.2Improved
add constrained decodingdiffusion LLMs · C++, JSON, SMILESfunctional correctness: +1.9% (Con.-), +2.2% (Con.)Improved
add constrained decodingDream 7B · JSONfunctional correctness: +6.9%Improved
add syntactic constraintsdiffusion LLMs · SMILESfunctional correctness: +0.2% average, from 1.5% baselineNo clear effect
increase number of infilling regions (runtime cost of constraining)constrained decoding · HumanEval C++ MRIruntime overhead: 67% -> 205%Hurt
model size effect on constraint overheadconstrained decoding · HumanEval C++ FIM/MRIruntime overhead: 320% (1.3B), 100% (7B), 20% (33B)Mixed
add constrained decodingdiffusion LLMs · C++, JSON, SMILESruntime: median overhead 30%Mixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy