.png){kind=icon}
Return
Two-stage distributionally robust decoding for reliability-oriented natural language generation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.ipm.2026.105000.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Reliability in natural language generation (NLG) is often compromised by semantically plausible but unsupported continuations. We present Two-Stage Distributionally Robust Decoding (TS-DRD), an inference-time framework that combines Wasserstein-robust logit correction with uncertainty-conditioned affine steering. The robust correction is analyzed at the level of the surrogate objective used inside the transport layer. A first-order variational characterization explains why the correction smooths semantically vulnerable token configurations, although it provides no theorem-level guarantee on factual correctness or hallucination rate. Factuality-oriented gains are therefore reported as empirical findings on public benchmarks. We evaluate TS-DRD on TruthfulQA (N = 817), HaluEval (N = 10,000), CNN/DailyMail (N = 11,490), and PopQA. Empirically, TS-DRD improves TruthfulQA MC2 accuracy from 53.8% to 58.2% over static robust baselines and reduces the HaluEval summarization hallucination rate from 28.8% to 24.2%. Disaggregated diagnostics further characterize the source of these gains: a baseline-correct vs. baseline-hallucinating bucket split shows that gains come predominantly from correcting hallucinated cases rather than from uniformly shortening responses; a held-out fidelity check between our single-pass uncertainty predictor and the offline Monte Carlo (MC) Dropout statistics it approximates yields Pearson r = 0.86; and per-method generation-length and entity-density measurements rule out trivial conservative shortening. These results support TS-DRD as a practical approach for reliability-oriented decoding in information processing tasks.
Keywords:
Decoding
Hallucination
Reliability
Robustness
Uncertainty
Wasserstein
Journal
I
IF:
6.9
Papers:
577
Citations:
0
Organization
No organization information available
Cited Papers
No cited papers available