The Flow has not summarised this story yet — read it at arXiv Computation and Language.
Depth pruning removes entire Transformer blocks to shrink and speed up language models, but the resulting damage is uneven across capabilities. In a study of four 1.2–1.7 B instruction‑tuned models, function calling and mathematical reasoning collapsed in every case, while other skills varied in order of degradation. Fine‑tuning a short ‘healing’ run can restore some abilities, with the choice of which blocks to retrain and the allocation of a limited token budget influencing recovery outcomes.
The study compares two delivery architectures for a multi‑turn agent that uses a tool to retrieve structured records. In the SYNTHESIS architecture, all records are fed into the model’s context, leading to a dramatic drop in accuracy as the table grows beyond about 74 rows. The BYPASS architecture keeps records in session state and delivers them via a typed side channel, maintaining high accuracy (92.3%) regardless of table size across four large language models.
arXiv:2605.19309v4 Announce Type: replace Abstract: Document Layout Analysis (DLA) pipelines provide structured page representations for retrieval-augmented generation, long-document question answeri...
arXiv:2605.26037v2 Announce Type: replace Abstract: We test the standard RLVR tool-use recipe -- GRPO on Qwen2.5-7B-Instruct -- on a deliberately minimal knowledge-graph tool API: four Freebase navig...
WeVisDoc is a two‑stage data‑centric framework designed to improve end‑to‑end document parsing. Stage I expands coverage by adding heterogeneous data and applying structure‑preserving degradation synthesis, while Stage II evaluates residual errors with a held‑out probe and uses those diagnostics to target data construction and token budget reallocation. The resulting WeVisDoc‑4B model achieves an overall score of 95.38 on OmniDocBench v1.6 and outperforms competing parsers across all evaluated settings, with Stage II delivering notable gains on degraded tracks.
The study examines how context projection—replacing older tool observations with concise, addressable excerpts—affects performance in ReVerPi, a Pi extension that archives observations and matches full to projected continuations. Across 86 source‑reading runs and 641 model requests, 15 completed pairs achieved identical success rates (12/15 per arm), while 12 boundary runs halted when the first arm failed, revealing that projection can reduce logical tokens by 25% but increase median pair tokens by 29% and total suffix requests from 35 to 55. The analysis highlights the importance of retaining all intervention boundaries, executing both arms independently, and reporting completion, token usage, and interaction details to accurately assess stopping rules and resource aggregation.