When Deep Research Agents Stagnate: Enhancing Reasoning with Retrieval-Aware Agent Control
Heydar Soudani, Elizabeth Lingg, Faegheh Hasibi, Navid Rekabsaz
cs.IR
2026-08-15
A bolt-on retrieval-aware controller cuts deep research agents' search calls by 14 on average while gaining up to 10 accuracy points across seven agents and three benchmarks.
Deep Research Agents, the tool-using agents behind products like OpenAI's Deep Research, run long loops of search and reasoning. This paper takes seven of them apart iteration by iteration (ReAct, GLM-4.7, GPT-oss-20b, Tongyi-DR, AgentCPM-Explore, WebWeaver, AgentCPM-Report, spanning four architecture families) and measures intermediate recall and accuracy at every step. The pattern is consistent: performance saturates after the first few searches, and most later iterations add nothing. Agents retrieve fewer and fewer new documents while revisiting pages they have already seen. In one traced example the answer was available at iteration 8; the agent ran 23. The root cause is that agents have no sense of their own trajectory: they do not know what they have covered, when to change direction, or when to stop. The authors call this reasoning stagnation.
RAAC (Retrieval-Aware Agent Controller) is a bolt-on controller that leaves the agent untouched. At each iteration it computes four unsupervised signals:
The controller maps these to one of three actions: continue; intervene, where a critical re-thinker LLM critiques why the trajectory is stale and issues a substantially different query; or stop, forcing the final answer. Signals use Qwen3-Embedding-4B; the controller and re-thinker use Claude Sonnet 4.5. Nothing is trained.
Across BrowseComp-Plus, NeuCLIR24, and Thomson Reuters' LegalSearch, adding RAAC never hurts and usually helps significantly, while cutting search calls sharply.
| Agent (base → +RAAC) | Metric | BrowseComp-Plus |
| Tongyi-DR | Accuracy | 37.9 → 47.9 |
| Tongyi-DR | Recall | 48.4 → 59.0 |
| WebWeaver | Recall / Acc | 64.9→67.2 / 47.8→51.3 |
| WebWeaver | Search calls | 75.0 → 50.7 (-33%) |
| Tongyi-DR | Search calls (NeuCLIR) | 62.1 → 34.4 (-45%) |
| All agents | Iterations (LegalSearch) | down 58% on average |
On BrowseComp-Plus the average accuracy gain is 3 points, peaking at 10 (Tongyi-DR), with 14 fewer search calls on average. The one exception confirms the mechanism: AgentCPM-Report tends to quit too early, so RAAC makes it search more, and recall jumps from 18.6 to 30.5. Ablations show Nov is the most effective primary signal; Cov helps only where human-written criteria exist. The controller adds about 10% latency per query, set against roughly 33% fewer searches.
Any team running a deep research product can borrow this directly: stagnation is architecture-independent, and a training-free external controller cuts cost and latency while improving accuracy. The win/loss analysis of controller actions (35% wins vs 18% losses in the stopped group) shows interventions redirect trajectories rather than interrupt blindly.
The authors list three: the controller is a hand-crafted prompt mapping signals to three discrete actions with fixed weights; only four signals are tested, leaving document content, retrieval confidence, and inter-document redundancy untried; the action set is only continue, intervene, and stop. Two more concerns from reading it: the critical re-thinker is itself an LLM, and the paper offers only indirect evidence that one LLM can reliably judge another's stagnation; LegalSearch is a closed Thomson Reuters dataset (the first author interned there), so those results cannot be reproduced externally.