Autoregressive ArcaneQA skews to seen data while discriminative Pangu generalizes
ysu_nlp · x · 2026-09-24
Follow-up evidence for the ACL'23 argument: autoregressive ArcaneQA's output distribution skews heavily toward seen data, causing distributional mismatch and generalization failure from seen to unseen, while discriminative Pangu shows much better-aligned output probability distributions over seen and unseen inputs.
More from Research
- Same model, two placements: agentic control completes 18/20 LEGO tasks vs 6/20 for code-as-policy — paigeinsf · 2026-09-24
- Gaia paper: metagenomic discovery with late-2024 LLMs predates Claude's find — owl_posting · 2026-09-24
- Yoav Goldberg: LLM Spotted the Pattern by Analogizing It to CRISPR — yoavgo · 2026-09-24
- TRACES: A New Benchmark That Grades AI Problem-Solving Process, Not Just Correct Answers — dr_cintas · 2026-09-24
- ECCV MMBU Benchmark Shows VLMs Answer Biomedical Questions Without Knowing What They See — davidjhwu · 2026-09-24
- Study: authors can reliably predict which of their papers will be highly cited — RexDouglass · 2026-09-24