ABSeeker goes after the quiet failure mode of every search agent I’ve had to train: the reward is binary and lands only at the very end, so the model never learns which of its twelve retrieval hops actually mattered. Its Answer-Backtracked Credit Assignment is clever precisely because it cheats with information you usually already have sitting in an eval set — the ground-truth answer — and uses it to score the path, not just the destination.
- 🎯 Backtrack from the answer to recover the intermediate clues a correct trajectory had to surface — no per-step human relevance labels
- 🔍 Score each step against those clues, converting one sparse outcome reward into dense step-level signal
- ⚡ Credit useful actions inside failed runs and suppress redundant ones, instead of discarding whole trajectories
- 📊 Feeds both SFT and RL — ABC-SFT reweights per-turn loss, ABC-GRPO uses the step scores as rewards
- ✅ Offline-friendly: it runs on trajectories you replay against known answers, so it slots into an existing harness before you stand up a live RL loop
- 💡 A 4B model on 8.5k examples reaches 37–39% on BrowseComp and, with context management, matches roughly 30B-scale agents
What I keep circling on from the HF paper page: clue recovery assumes the answer uniquely implies the clues. For obscure multi-hop trivia that holds, but for the fuzzy enterprise questions my retrieval stack actually fields — where three different evidence paths all reach a defensible answer — does anchoring on one backtracked clue set quietly punish the agent for taking a different-but-valid route?