Paper finds decoding collapse behind failed self-distillation in search agents
_reachsumit · x · 2026-07-21
A new paper asks why feedback-augmented self-distillation fails to improve retrieval-interleaved search agents.
The authors identify decoding collapse as the core failure mode and show that EMA teacher regularization can help. The post frames the result as a diagnosis of why self-distillation does not reliably improve these agents, rather than a simple benchmark win.
More from Research
- LFM2 tokenizer expansion cuts Thai tokens 4× and speeds on-device decoding up to 3.7× — maximelabonne · 2026-07-21
- Argus improves indoor panoramic 3D reconstruction with covisibility and geometry transformers — ducha_aiki · 2026-07-21
- WAIC robots are now hitting commercially useful success rates, says a recap — chris_j_paxton · 2026-07-21
- Unitree launches a remote real-robot benchmark on its own G1 fleet — chris_j_paxton · 2026-07-21
- Adding order metadata makes VLM error detection collapse, new benchmark shows — m_wulfmeier · 2026-07-21
- New prompt template aims to improve spatial reasoning and cut model laziness — legit_api · 2026-07-21