Long-horizon research tasks break when a single policy must plan, query, filter evidence and synthesize answers while accumulating ever-growing noisy context. The core insight of IterSynth is that separating ‘‘where to search’’ from ‘‘how to incorporate findings’’ and keeping a compact evolving summary as the search state reduces capability coupling and context bloat, improving both exploration and memory quality.
Key Findings
- Role decoupling: Alternating a Planner (decides next query or final answer) with a Synthesizer (updates the global summary from retrieved passages) lets one shared LLM specialize behaviorally without extra model parameters. This separation reduces premature termination and redundant searches common in ReAct-style single-policy agents.
- Summary-as-state: Treating the summary as the persistent, up-to-date memory (instead of an ever-growing linear trace) limits context noise and makes evidence retrieval and reuse more reliable across many iterations.
- RDPO training: Role‑Decoupled Policy Optimization combines terminal outcome rewards with turn-level rubric evaluations and computes role-specific advantages, yielding more precise credit assignment and better performance than outcome-only RL.
- Empirical gains: IterSynth-8B achieves an average score of 50.7 across five long-horizon benchmarks, improving over prior ≤8B agents and providing consistent zero-shot prompting gains for frontier models.
Who it's for and tradeoffs
Great fit if you need to build or evaluate LLM-driven research/long-horizon search agents that must make many iterative tool calls and maintain a compact, reliable memory. IterSynth is also useful as a prompting paradigm for improving proprietary large models' zero-shot deep-search behavior. Look elsewhere if you require a simple single-step QA system (IterSynth targets multi-step evidence-gathering workflows), if you cannot run iterative tool interactions, or if latency/engineering constraints make repeated retrieval-and-summarize loops impractical.
Where it fits
IterSynth sits between single-trajectory ReAct agents and full multi-agent systems: it achieves role specialization without multiple models and focuses on improving long-horizon exploration, evidence integration, and memory control. It is most directly comparable to other iterative summary-based frameworks for deep research agents.
How it works (brief)
At each iteration the Planner observes the original question plus the current global summary and either issues a search query or terminates with an answer. A retrieval step obtains passages for that query. The Synthesizer then ingests the retrieved evidence and updates the global summary by extracting findings, filtering irrelevant content, and resolving local inconsistencies. Training proceeds from supervised SFT trajectories to RDPO, which optimizes terminal and turn-level rubric rewards with role-specific advantage estimation.