Source-linked AI summary
G\"odel Forest: Balancing Search Depth and Breadth for Data-Centric Recursive Self-Improvement
Ziqi Zhao, Fanqing Meng, Haocheng Lu, Lingxiao Du, Qiguang Chen, Mengkang Hu, Xiao-Ming Wu
TL;DR
Data-centric recursive self-improvement updates model parameters through agent-generated data, but expensive validation makes deep search and broad exploration difficult to reconcile. Gödel Forest coordinates persistent search trees with asynchronously shared lessons; across six RSIBench-Data benchmarks, it outperforms the single-agent baseline by an average of 10.70%.
Problem
Expensive training-in-the-loop validation leaves data-centric RSI balancing deep hypothesis pursuit against broad exploration, while single-agent and parallel approaches each have search or coordination drawbacks.
Source: §1 · p. 2, §3.1 · p. 4Method
Gödel Forest combines persistent per-agent search trees with shared memory that asynchronously distributes compact empirical lessons.
Source: §1 · p. 2, §3 · p. 3, §3.3 · p. 6Results
Across six RSIBench-Data benchmarks, Gödel Forest gains an average 10.70% over the single-agent baseline and reduces wall-clock time on five of six benchmarks.
Source: §1 · p. 3Takeaways & Limitations
The reported +7.00% gain over independent parallel agents supports shared-memory co-evolution as a useful coordination component in this benchmark setting.
Source: §5 · p. 10
Abstract
from arXiv · showhide
Recursive self-improvement (RSI) aims to achieve compounding gains by having models improve themselves. While most existing RSI systems optimize external agent harnesses or prompts around a frozen base model, data-centric RSI directly updates the model's own parameters by training on agent-generated data. However, because validating data strategies requires expensive model training, existing methods face a fundamental dilemma: a single agent gets trapped in narrow directions and lacks exploration breadth, while naive parallel search or heavy trace sharing sacrifices long-horizon search depth. To address this challenge, we introduce G"odel Forest, a multi-agent framework that organizes recursive self-improvement as an ensemble of co-evolving search trees. In G"odel Forest, each agent autonomously grows a persistent tree, deepening, branching, or pruning data strategies based on model feedback to secure depth, while parallel trees explore distinct regions of the data space to expand breadth. Crucially, rather than leaving trees isolated or flooding them with heavy execution logs, a dynamically co-evolving memory connects the forest: agents continuously distill their successes and failures into compact procedural lessons anchored to a global leaderboard. Through this forest ecosystem, a dead-end in one tree instantly warns the whole forest against unpromising paths, while an empirical breakthrough quickly seeds new exploration branches in neighboring trees. Evaluated on RSIBench-Data across six diverse domains, G"odel Forest outperforms the single-agent baseline by an average of 10.70% while reducing wall-clock time on five tasks. Ablations confirm that co-evolving shared memory yields a +7.00% gain over independent parallel search, demonstrating that collective distillation is key to scalable self-improvement. The code is available at https://github.com/evolvent-ai/Godel-Forest.
1. Introduction
Data-centric RSI trains candidate models on agent-generated data, but costly training-in-the-loop validation makes deep search and broad exploration difficult to combine. Gödel Forest uses persistent search trees and shared memory to address this trade-off, and it outperforms a single-agent baseline across six benchmarks.
1. Introduction
- Data-centric RSI validates candidate data strategies through model training and downstream evaluation, making exploration expensive and leaving single-agent depth and parallel breadth in tension.Single agents can follow narrow trajectories; parallel approaches may repeat failures or suffer from verbose trace exchange.Source: §1 · p. 2
- Gödel Forest combines autonomous persistent search trees with asynchronously shared, compact lessons that carry failures and breakthroughs across the ensemble.A global leaderboard anchors the lessons; negative findings warn against dead ends, while successes seed new branches.Source: §1 · p. 2
- Gödel Forest beats the single-agent baseline by an average of 10.70% across six benchmarks and reduces wall-clock time on five of six.The introduction also reports SWE-bench Verified rising from 35.00% to 54.00%.Source: §1 · pp. 2–3
- The paper frames data-centric RSI as open-ended data-evolution search and evaluates the framework across six diverse RSIBench-Data benchmarks.Its stated contributions include gains over single-agent and independent-parallel baselines, improved wall-clock efficiency, and advantages from shared-memory co-evolution.Source: §1 · p. 3
2. Related Work
Prior empirical Gödel systems evolve external harnesses or prompts, while data-centric RSI updates model parameters through training on agent-generated data. Gödel Forest combines parameter-updating search with asynchronous shared-memory coordination.
Related work
- Empirical Gödel systems replace formal proof with observed task utility but restrict self-modification to external harnesses or prompt wrappers around frozen models.Sharing verbose execution traces can also cause context bloat in open-ended research spaces.Source: §2 · p. 3
- Data-centric RSI updates model parameters, but existing systems often rely on single-agent search or uncoordinated parallel sampling.The passage describes localized data-distribution search and redundant failures under constrained budgets.Source: §2 · p. 3
- Gödel Forest presents asynchronous memory co-evolution as coordination between private trees for deep refinement and shared lessons that reduce redundant exploration and propagate recipes.The framework is grounded in real parameter updates.Source: §2 · p. 3
3. Gödel Forest
Gödel Forest couples persistent local search trees with asynchronous shared memory to pursue deep data-strategy refinement and broad exploration. It evaluates candidate data programs through trained target models, while shared outcomes and lessons coordinate workers.
3. Gödel Forest
- The framework organizes concurrent workers as persistent search trees coupled through dynamically shared evolution memory.The shared memory is a non-blocking public blackboard.Source: §3 · p. 3, §3.3 · p. 5
3.1. Problem Formulation and Single-Tree Foundations
- Data-centric RSI treats each archived agent as a data-engineering program that generates training data, trains a candidate from fixed initial parameters, and evaluates downstream utility.Earlier models' empirical performance and failure diagnostics update the agent's persistent state and guide later data decisions.Source: §3.1 · p. 4
- Full-loop training and evaluation make data-hypothesis validation costly, so a single tree risks narrow path-dependent search while parallel search may repeat failures or incur trace-sharing overhead.Reconciling deep multi-round pursuit with broad, lightweight coordination is the stated objective.Source: §3.1 · p. 4
3.2. Persistent Tree Search: Local Deep Hypothesis Pursuit
- Workers diagnose prior failure modes, propose successor data programs from private trajectories and shared memory, and deepen, revise, or backtrack their private trees.Candidates first receive a lightweight diagnostic-subset evaluation and proceed to remaining tasks only if they pass the gating threshold; otherwise evaluation ends early.Source: §3.2 · p. 4
3.3. Forest Co-Evolution via Shared Evolution Memory
- A shared leaderboard aggregates verified utilities across workers, while structured lessons record tested hypotheses, utility shifts, outcome attributions, and actionable rules.Lessons replace raw samples or execution traces as the coordination unit.Source: §3.3 · pp. 5–6
- Workers asynchronously retrieve relevant lessons, use failures as guardrails and successes as seeds, then publish new lessons to the shared memory.The protocol is pull-based and non-blocking, without global synchronization barriers.Source: §3.3 · p. 6
4. Experiments
Gödel Forest was evaluated across six benchmarks, with gains over the single-agent baseline and lower wall-clock time on five. Cross-backbone tests, ablations, and search trajectories examine its generality, coordination, and search dynamics.
4.1. Experimental Setup
- Evaluations cover six task domains, using Qwen3.5-35B-A3B-Base candidates, LoRA SFT, and a fixed 16-hour and $500 forest budget.The primary baseline is single-agent Claude Code with Sonnet-5; selected checkpoints receive official post-search evaluation.Source: §4.1 · p. 6
4.2. Main Results
- Across all six benchmarks, Gödel Forest consistently outperformed the single-agent baseline, averaging a 10.70% absolute gain; on SWE-bench Verified, performance rose from 35.00% to 54.00% (+19.00%).Source: §4.2 · p. 7
- Lower wall-clock time on five benchmarks accompanies higher expenditure on long-horizon tasks, where evolved models sustain longer successful interactions.The paper attributes faster completion to asynchronous execution and shared-memory pruning of unviable branches.Source: §4.2 · p. 7
4.3. Cross-Backbone Generalization: Results with Codex
- With Codex workers, official scores rose from 33.00% to 44.00% on SWE-bench Verified and from 20.22% to 23.60% on Terminal-Bench 2.0.Wall-clock time was 6.81h versus 4.41h and 9.85h versus 9.56h, respectively.Source: §4.3 · p. 8, Figure 2 · p. 8
4.4. Ablation Study: Dissecting Exploration Breadth and Coordination
- 45.00% to 54.00% performance as workers scale from K=1 to K=3; K=4 plateaus at 54.00% while using 15.80h and $404.43.The study identifies K=3 as the Pareto balance between breadth, depth, and resource use.Source: §4.4 · pp. 8–9, Figure 3 · p. 8
- Three independent workers without shared memory reach 47.00%, a 7.00% drop versus Gödel Forest, despite comparable budget use.The paper reports repeated spending on dead ends already diagnosed by peers.Source: Figure 3 · p. 8, §4.4 · p. 9
4.5. Search Dynamics and Cross-Worker Knowledge Synthesis
- Search trajectories are non-monotonic; parallel trees can advance complementary branches while a persistent archive preserves prior performance peaks.The figure marks diagnostic and full evaluations, with selected checkpoints shown as stars.Source: §4.5 · p. 9, Figure 4 · p. 9
- Workers distilled compact lessons instead of sharing raw data or verbose logs: final-turn masking curbed runaway generation, and supervised verification actions yielded the selected 55% checkpoint.The final-turn lesson was adopted by another worker, whose score rose from near-zero to 22%.Source: §4.5 · pp. 9–10
5. Conclusion
Gödel Forest addresses the depth-versus-breadth search dilemma through persistent local trees and asynchronous shared memory. Across six RSIBench-Data benchmarks, it reports gains over single-agent and parallel baselines with lower wall-clock latency.
5. Conclusion
- +10.70% average score over single-agent and parallel baselines, with a +7.00% ablation gain and lower wall-clock latency across six RSIBench-Data benchmarks.The conclusion presents these results as validating asynchronous tree co-evolution as a scalable self-improvement paradigm.Source: §5 · p. 10
- Persistent local search trees use diagnostic gating for deep hypothesis pursuit, while parallel trees explore distinct data regions.The trees coordinate asynchronously through shared memory.Source: §5 · p. 10
- Compact shared lessons anchored to a leaderboard let failures prune dead ends and breakthroughs seed new branches across the forest.This creates a collective feedback loop from individual tree outcomes.Source: §5 · p. 10