UCL study finds LLMs share a Fisher-Rao information geometry that enables low-distortion model interventions
UniversityCollegeLondon · hf · 2026-09-23
Researchers from University College London published a study arguing that the Fisher-Rao information geometry of next-token probabilities is shared across LLM architectures, making model behavior linked, learnable, and controllable.
- Cross-architecture共性 is replaced by: Shared structure: geometry links behavior across different models
- Intervention: enables low-distortion, targeted interventions on model behavior
- Dynamics: predicts learning dynamics and allows control transfer across models
More from Research
- Steering VLA robot models at inference time lifts grasp rate from 14% to 74% without retraining — drmapavone · 2026-09-23
- Dev strips decoding from Qwen 3.5 9B to serve token probabilities locally at 100ms — telltaleatheist · 2026-09-23
- Microsoft's Agensh Scales Multi-Agent Systems to 1,024 Agents Without a Central Orchestrator, Boosting Test-Pass Rate to 55% — andrew_n_carr · 2026-09-23
- ValsAI says 10 Claude Opus 5.5 agents built a Lean-verified algorithm improvement in 15 hours — alejandroll10 · 2026-09-23
- Stanford-led Terminal-Bench-Science debuts: top model solves just 30% of 70 expert tasks — geoffwolfe · 2026-09-23
- Untrained Qwen beats purpose-trained decision model by 35 points in real repo test — KingPinX · 2026-09-23