MARS trains RL agents to search raw corpora via shell commands, no retriever needed
_reachsumit · x · 2026-11-10
MARS explores retriever-free multilingual RAG: RL-trained agents interact with raw corpora via executable shell commands, avoiding learned retrievers' cross-lingual errors and indexing overhead.
Challenges include sensitivity to entity surface forms and weak outcome-only supervision for multi-turn search. MARS introduces corpus-native entity grounding to recover reliable lexical anchors and a redundancy-aware search reward to keep trajectories concise and stabilize RL training. It consistently beats strong retrieval-based and direct-corpus-interaction baselines, with robust cross-lingual generalization and far fewer redundant interactions.
More from coding & agent
- Dev Asks Reddit Whether Ramp Inspect Survives Legacy Codebases With Undocumented Quirks — InternationalKick296 · 2026-11-10
- Claude Duet cleared her inbox and did the tasks — with a git repo as process memory — RealGeneKim · 2026-11-10
- Daiku hands your logged-in browser to AI agents — and it politely declined a troll order — TejasKumar_ · 2026-11-10
- Browser use: Muse Ember matches GPT-6 Luna at 3 cents per task — WenhuChen · 2026-11-10
- Model self-postmortem: four architecture choices that made my Three.js Minecraft lose to Fable 5.2 — teortaxesTex · 2026-11-10
- Reviewing 200-line PRs by day, AI writes 18,000 lines in 4 hours by night — facontidavide · 2026-11-10