23 papers this week in Harnesses & orchestration12 active threadsbusiest: Agent Security & Attacks

Harnesses & orchestration research

Small language model matches larger ones on conversational tasks with half the memory

NVIDIA introduces Nemotron 3.5 Bolt E4B 4GB, an open small language model optimized for fast conversational agents in local settings. Built from Gemma-4-E4B-it and further trained with reinforcement learning, the model achieves the highest average benchmark score across 21 tasks, spanning multi-turn tool use, instruction following, and input grounding, while generating fewer output tokens than Gemma-4-E4B-it and Qwen3.5-4B in both reasoning and non-reasoning modes.

4 days ago
Agents

Stopping agents when further gains are unlikely saves token costs without hurting quality

The authors introduce GainCtrl, a method that uses a lightweight CPU regressor to predict the remaining score gain if an LLM agent continues its current attempt. By stopping attempts when the predicted gain is low, the controller reduces token consumption while preserving the quality of the best result across attempts. Across three tool-using environments and three different LLM actors, GainCtrl consistently achieves better token–quality trade-offs than full execution and a prior early-stopping baseline.

4 days ago
Code & software

Coding agents learn more when training projects are revised automatically

Tian and colleagues built an automated pipeline that turns real software repositories into training tasks where coding agents must construct complete projects from written specifications. Training on 10,159 resulting environments improved a language model’s coding performance, with additional gains when the pipeline revised tasks using automated measures of validity, implementation scope and connections between components.

4 days ago