Salesforce AI's SRD aligns agent foresight with hindsight at zero test-time cost
lupantech · x · 2026-10-21
Salesforce AI Research, with UCSD, Stanford and others, introduces Self-Retrospection Distillation (SRD), teaching agents to learn what they should have seen coming.
Key points:
- Prospective learning: the agent forms a Foresight prediction before acting; after acting, Hindsight is distilled from traces and the two are aligned via divergence D(pfore || phind)
- Unlike RLVR (one scalar reward per trajectory, e.g. GRPO/DAPO) and retrospective methods (OPSD/SkillRL, where hindsight is only fed back as context), SRD distills hindsight directly into foresight
- Learns from every rollout, including all-fail groups
- Zero extra test-time inference cost
Open source: github.com/SalesforceAIResearch/SRD
More from coding & agent
- Running >1 autonomous agent: doc bloat, review bottlenecks, and why "don't edit" fails — menhguin · 2026-10-21
- LangChain expands agent redteaming beta: 34 hypotheses tested, 9 findings in one run — hwchase17 · 2026-10-21
- OvertChat: a polished self-hosted ChatGPT-style UI for local models, one-command install — anitamaxwynnn69 · 2026-10-21
- Smart Connections demo: Obsidian surfaces past decisions to reshape your plan — AINewsletter · 2026-10-21
- Ampersand v0.1 Ships via Agent API: Hosted Sessions, OpenAI-Client Compatible — gharik · 2026-10-21
- Pi Durable: Mario Zechner and Armin Ronacher on crash-safe long-running agents — mitsuhiko · 2026-10-21