← Knowledge Atlas · Entity
Maximilian Schall
Maximilian Schall：NLP 研究者，轨迹偏差概念命名者（RANLP 2025），约束解码语义代价研究
实
Maximilian Schall
NLP researcher — with Gerard de Melo, the first to systematically demonstrate the semantic cost of constrained decoding
Schall, with de Melo, named and empirically established “trajectory bias” through cross-benchmark experiments on 11 models — constrained decoding preserves syntactic compliance but systematically harms semantic correctness. This is the first large-scale quantitative study of the cost of constrained decoding.
Key Works
2025The Hidden Cost of Structure (RANLP 2025)First systematic demonstration of the syntactic-compliance vs. semantic-correctness trade-off in constrained decoding — 11 models, cross-benchmark experiments
2025Draft-Conditioned Constrained Decoding (arXiv:2603.03305)With Castillo & de Melo — a two-stage separation (draft generation + format constraint) as an algorithmic fix for trajectory bias
Core Findings Base vs. Instruction-Tuned Models The two classes diverge sharply under constraints — instruction-tuned models are more fragile, with larger semantic damage Draft-Conditioned Solution Generate a draft freely first (preserving semantic integrity), then format it with constrained decoding — decoupling semantic generation from format enforcement Theoretical Implication Constrained decoding is not a “free lunch” — the price of format compliance is a measurable semantic bias
→ Trajectory Bias · Constrained Decoding · Structured OutputsRANLP 2025 / arXiv:2603.03305
Maximilian Schall
简介
Maximilian Schall，NLP 研究者，与 Gerard de Melo 合作研究 LLM 约束解码与推理质量。
主要贡献
- 轨迹偏差（Trajectory Bias）命名与实证（RANLP 2025）：首次系统性证明约束解码在保持句法合规的同时会损害语义正确性，通过对 11 个模型的跨基准实验识别了基础模型与指令微调模型在约束下的分歧行为。
- Draft-Conditioned Constrained Decoding（arXiv:2603.03305, 2025）：与 Castillo、de Melo 合作提出通过草稿-格式两阶段分离来缓解轨迹偏差的算法解。
相关概念
References
- Schall & de Melo. “The Hidden Cost of Structure.” RANLP 2025.
sources/ranlp-2025-hidden-cost-constrained-decoding.md