Research map / Probabilistic and scientific modeling / Variational and generative sampling
Mitigating Bias in Locally Constrained Decoding via Tractable Proposals
Meihua Dang, Linxin Song, Honghua Zhang, Jieyu Zhao, Guy Van den Broeck, Stefano Ermon
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 changed | Where | Measured | Effect |
|---|---|---|---|
| mask with global feasibility within token budget (GCD) instead of local masking (LCD) | SMC proposal · xLAM JSON, k=4 | constraint satisfaction rate (%): 95.26 -> 100.0 | Improved |
| mask with global feasibility within token budget (GCD) instead of local masking (LCD) | SMC proposal · xLAM (JSON and Python-like) | accuracy | Improved |
| add HMM probabilistic information (P-GCD) on top of GCD | SMC proposal · xLAM | convergence speed | Improved |
| use P-GCD (H=16384) instead of LCD proposal | SMC on xLAM JSON · xLAM JSON, k=16 | accuracy (%): 69.7 -> 71.6 | Improved |
| use P-GCD with single particle | SMC on xLAM JSON · xLAM JSON, k=1 | accuracy (%): 66.4 -> 66.5 | No clear effect |
| increase HMM hidden size from 1024 to 16384 (ablation) | P-GCD · xLAM JSON | accuracy (%): k=2 68.6 -> 69.2; k=16 71.5 -> 71.6 | Improved |
| increase HMM hidden size to 16384 (ablation) | P-GCD · xLAM latency | ms/token: 5.26 -> 17.30 | Hurt |
| use GCD and P-GCD instead of LCD on Spider | SMC proposal · Spider, k=8 | execution accuracy (%): 67.2 -> 69.0 (GCD), 69.1 (P-GCD) | Improved |
| use P-GCD proposal with a single particle on Spider | SMC proposal · Spider, k=1 | execution accuracy (%): 59.8 -> 56.2 | Hurt |
| add P-GCD potential to P-GCD proposal (ablation) | SMC on CommonGen · CommonGen (GPT2-Large) | BLEU-4 | Improved |
| use NFAs instead of DFAs for SQL constraint | constraint automaton · Spider SQL grammar, character level | number of states: 57,263 -> 7,999 | Improved |
| use tensorized GCD instead of unconstrained decoding | decoding · xLAM latency | ms/token: 1.69 -> 1.86 | No clear effect |
Closest papers by meaning
- Constrained Decoding for Diffusion Language Models via Efficient Inference over Finite Automata (NeurIPS 2026 · similarity 0.91)
- Grammar-Aligned Decoding (NeurIPS 2024 · similarity 0.88)
- Controlled LLM Decoding via Discrete Auto-regressive Biasing (ICLR 2025 · similarity 0.87)
- Constrained Sampling for Language Models Should Be Easy: An MCMC Perspective (NeurIPS 2025 · similarity 0.87)
- A Theoretical Perspective for Speculative Decoding Algorithm (NeurIPS 2024 · similarity 0.86)
- Cactus: Accelerating Auto-Regressive Decoding with Constrained Acceptance Speculative Sampling (ICLR 2026 · similarity 0.86)
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.
