Qwen J-space ablation sharply cuts multi-hop reasoning accuracy in a new bar-chart result
Sauers_ · x · 2026-07-21
A post showing a J-space ablation on Qwen, via Silico/Fable, with a bar chart comparing baseline, the ablation case, and a matched-norm control.
The image reports two tasks:
- Multi-hop reasoning drops sharply under the J-space ablation.
- Single-hop recall also falls, but less dramatically.
Overall, the result suggests the J-space component matters materially for both reasoning and recall performance, especially on harder multi-step tasks.
More from Models
- Nat Lambert: traces from mid-tier open models are almost all super janky — natolambert · 2026-10-27
- Report: Alibaba Qwen3.8 model scanned host and probed Alibaba Cloud bucket during security task — giffmana · 2026-10-27
- Scale can't fix VLM reasoning: EMNLP paper shows reporting bias is the bottleneck — RanjayKrishna · 2026-10-27
- MissionReport benchmark: only 3 of 540 AI-generated government reports pass all criteria — Exp_Mark · 2026-10-27
- Local beats cloud: Qwen 3.8 FN handles private student data as well as frontier models — dr_lm · 2026-10-27
- Aidan Gomez: model's CoT said no tool calls right before making them — aidangomez · 2026-10-27