News · 2026-08-08

A 4B search agent matches 30B by grading its own failed attempts

A 4-billion-parameter web-search agent trained on 8,500 examples now matches models roughly seven times its size on the hardest open web-research benchmark. The technique has nothing to do with scale: it works backwards from the known correct answer to reconstruct the intermediate clues a searcher would have needed, then grades each individual step of a search against those clues -- so a useful step buried inside a run that ended wrong still earns credit.

Key facts

• ABSeeker is built on a Qwen3.5-4B base and trained on 8,500 examples.
• It reaches 37.3% on BrowseComp and 39.1% on its Chinese counterpart; with context management, 55.3% and 52.9%.
• Those scores match agents of roughly 30 billion parameters -- about seven times the size.
• Primary source: ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment, with code released on GitHub.

BrowseComp is a test of finding facts that are genuinely hard to find on the live web -- the kind where the answer exists but is three obscure hops from any obvious query. Solving one takes a long chain of actions: search, read, discard, refine, cross-check, and eventually commit. That chain is what makes training these agents hard.

The standard approach rewards the whole chain by its ending. Right answer, every step gets a gold star; wrong answer, every step gets penalised. Anyone who has debugged something for three hours knows why this is a terrible teacher. Most of that time was well spent -- you eliminated possibilities, found the relevant subsystem, narrowed it down -- and then made one bad call at the end. Flat outcome scoring tells the model that entire session was worthless. It also tells it that the twenty wasted searches inside a lucky success were brilliant. This is the credit assignment problem, and in long-horizon agents it is the dominant source of training noise.

ABSeeker's move is to use information that is available at training time but not at inference time: the correct answer. Given a hard question and its verified answer, the system first runs what the authors call Answer-Backtracked Clue Recovery -- tracing backwards from the answer to reconstruct the intermediate facts a solver would have had to establish along the way. Then Clue-Anchored Step Scoring evaluates each step of an actual trajectory against that reconstructed set of clues. A step that surfaces a needed clue scores positively even if the run later collapsed. A step that abandons an established clue and reverts to a discarded candidate scores negatively even if the run happened to end correctly.

The analogy is exam marking. Grading only the final number tells a student nothing about which line of their working went wrong. A teacher who marks the working can say: those four steps were right, this fifth one is where you lost it. ABSeeker manufactures the mark scheme automatically, by reasoning backwards from the answer key.

Those dense per-step scores then feed two training stages. ABC-SFT uses them to reweight the loss on each conversational turn during supervised fine-tuning, so the model imitates good steps more strongly than bad ones. ABC-GRPO plugs them in as the reward signal for group relative policy optimization, the reinforcement-learning method behind most recent reasoning models. Both stages keep failed trajectories in the training mix instead of discarding them, which is where the sample efficiency comes from: 8,500 examples is a very small budget, and the reason it suffices is that nothing gets thrown away.

The efficiency claim is the part worth dwelling on. Matching a 30-billion-parameter agent with a 4-billion-parameter one is not a scaling result; it is evidence that a large fraction of what bigger agents were buying with parameters was the ability to absorb an extremely noisy training signal. Clean the signal and you need much less model. That is the same logic behind reinforcement learning with verifiable rewards generally, pushed one level finer -- from verifying the answer to verifying the path.

The constraint is real and the authors state it. The method needs a unique verifiable answer and an evidence chain you can walk backwards along. That describes web research, fact-finding, and multi-hop question answering. It does not describe open-ended dialogue, creative work, or tasks where several genuinely different routes are all correct. There is also a subtler risk: the recovered clues are a reconstruction, not a record. If the agent found a valid route the backtracker never considered, the scorer will mark good work as bad -- a way of quietly teaching a model to search the way the reconstruction expects rather than the way that works.

ABSeeker is not alone in this direction. The same arXiv batch carried AgentOPSD, which aggregates token-level teacher-student gaps into turn-level evidence and updates a belief about each turn's value in log-odds space, and EnvACE, which internalises environment dynamics through rehearsal. Different mechanisms, one shared conviction: a single number at the end of a long run is not enough supervision to train a competent agent, and the field has started building the missing signal rather than waiting for scale to paper over it. That echoes recent findings that an RL trainer can invent its own reward when the judge stays silent -- and the more of the reward a system manufactures, the more it matters whether the manufactured reward is measuring the right thing.

Primary source, verified: read the paper → (arXiv 2608.05102)

Ground Truth. (2026, August 8). A 4B search agent matches 30B by grading its own failed attempts. Ground Truth. https://groundtruth.day/news/a-4b-search-agent-matches-30b-by-grading-its-own-failed-attempts.html

@misc{groundtruth:a-4b-search-agent-matches-30b-by-grading-its-own-failed-attempts,
  title  = {A 4B search agent matches 30B by grading its own failed attempts},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/a-4b-search-agent-matches-30b-by-grading-its-own-failed-attempts.html}
}

Topics: agents · reinforcement-learning · research · search · credit-assignment