structured-output

structured-output · output-formatting · dream-7b, llada-8b

recommended confidence: moderate no check recorded

FA-constrained decoding substantially improves accuracy on schema-constrained generation

• On BFCL-Live, constrained decoding raises Dream-7B accuracy from 63.9% to 71.5% under greedy decoding and from 22.3% to 69.0% under stochastic sampling, with under 5% wall-clock overhead. arXiv:2607.07026
• The unconstrained baseline degrades sharply when the required output format switches from JSON to Python, and constrained decoding largely closes this format gap, indicating the failures are formatting-related rather than intent-related. arXiv:2607.07026

Constrained Decoding for Diffusion Language Models via Efficient Inference over Finite Automata · arXiv preprint, 2026 not read in full

structured-output · output-formatting · llama-3.1

recommended confidence: moderate no check recorded

GCD/P-GCD proposals lift constraint satisfaction from 95.26–99.20% (LCD) to 100.0% on structured JSON/Python function-calling outputs

• On xLAM function calling with Llama-3.1-8B-Instruct, using tensorized globally constrained decoding (GCD) or probabilistic GCD as the SMC proposal achieved 100.0% constraint satisfaction at k=4 and k=16 particles in both JSON and Python-like output formats, while standard locally constrained decoding reached only 95.26% (JSON, k=4), 95.40% (JSON, k=16), 98.12% (Python-like, k=4), and 99.20% (Python-like, k=16). arXiv:2606.01926
• Under the same SMC sampling setup on xLAM, (P-)GCD proposals achieved substantially higher accuracy than LCD proposals at the same number of particles for both JSON and Python-like output formats, converging to the rejection-sampling ground truth with significantly fewer particles. arXiv:2606.01926

Mitigating Bias in Locally Constrained Decoding via Tractable Proposals · arXiv preprint, 2026 not read in full