Daily Research Brief 2026-08-08

📊 Token usage: estimated from retrieval and writing scale.

Covers the latest AI research, open source and industry moves, updated daily.

Editor’s Note

Today’s eight papers converge on the same sentence: the agent bottleneck is not ‘model too weak’ but ‘signal too sparse, shell too unstable’. MERIT lifts Spider from 66.34% to 69.79% with a bipolar causal memory and zero parameter changes; AgentOPSD turns sparse outcome signals into dense training signals.

1. Latest arXiv Papers

1. Causal Episodic Memory for Feedback-Driven Agent Repair (MERIT) — https://arxiv.org/abs/2608.05906

2. Contextual Information Policy Optimization for Search Agents (CIPO) — https://arxiv.org/abs/2608.06128

3. AgentOPSD: Recursive Self-Distillation for Agentic Reinforcement Learning — https://arxiv.org/abs/2608.05987

4. Activity Frames: Deterministic Screen-Activity Compilation for Agent Memory and Replay — https://arxiv.org/abs/2608.05784

5. Learning Globally Reusable Skills for Coding Agents (GSE) — https://arxiv.org/abs/2608.06153

6. SkillTrace: Multi-Trace Provenance Auditing for LLM-Agent Skill Reuse — https://arxiv.org/abs/2608.05204

7. DreamGuard: Efficient Runtime Guardrail for LLM Agents via Risk-Aware World Model — https://arxiv.org/abs/2608.05695

8. Privileged, but Biased: How PI-Conditioned Teachers Break Self-Distillation — https://arxiv.org/abs/2608.04794