AI
More Choices, Fewer Decisions: Ordinal-Scale Bias in JEV-like Direct-Decision Models
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 29/09/2026, 20:00:00 UTC
Briefing created: 01/10/2026, 09:20:12 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 01/10/2026, 03:18:24 UTC.
Briefings
A study of JEV~1.13 and three open KEV models reveals a systematic compression of decision-stage candidate space, where models utilize only 67-76% of effective gold support on ordinal tasks compared to 87-102% on nominal tasks. This bias, termed ordinal scale-utilization bias, is learned via targeted BA-LoRA post-training, raising gold-relative utilization from ~47% to 86% on eight supervised scales. The research analyzes ANLI and 36 ordinal datasets, noting that while accuracy remains high, reliability suffers due to this structural compression.
Key points
- JEV~1.13 and three open KEV models utilize only 67-76% of effective gold support on ordinal tasks, whereas nominal tasks show 87-102% utilization.
- Targeted BA-LoRA post-training successfully raises gold-relative utilization from approximately 47% to 86% on eight supervised scales, proving the compression is learnable rather than an immutable architectural limit.
Context and limitations
The study covers 36 ordinal datasets and ANLI, with specific limitations on candidate order randomization and the scope of supervised scales for the post-training intervention.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 01/10/2026, 09:20:12 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeieb2krvwgphwpwnu7dvl2xcenwshywp5hnufuwitw5gntinaz2z3m