Berkeley team shows LLMs lose info via recursive self-summarization, fixes it with RL
berkeley_ai · x · 2026-11-05
A Berkeley team's TMLR paper examines how LLMs degrade when recursively self-summarizing in long-horizon information-gathering tasks, and argues outcome-reward RL alone is insufficient without expert demonstrations. Their fix: grade belief states by reconstructing information from the trajectory, using an LLM-based reconstruction reward (see diagram).
More from Research
- Cartridges: 'Self-Study is a Moving Target' — a label-free metric for KV-cache compression — simran_s_arora · 2026-11-05
- Baidu open-sources ERNIE 5.5 Flash: 280B-param MoE with 14B active and 1M context — teortaxesTex · 2026-11-05
- USC's WACL teaches robots whole-arm manipulation while the hand stays busy — ebiyik_ · 2026-11-05
- nanoGRPO: train your own tiny R1-style reasoner on an 8GB RTX 4060 — joey00072fp8 · 2026-11-05
- First-ever asymptotic for E.P. Stein numbers proved with an AI collaborator, published on arXiv — arampell · 2026-11-05
- Weight-space programming's soft claims rest on 3-4 correct answers on toy tasks — maier_ak · 2026-11-05