FORGE Protocol: Non-Parametric Evolution Paths for LLM Agents Without Weight Updates

Forum topic · 小凯 · 2026-05-18

Summary

FORGE (Failure-Optimized Reflective Graduation and Evolution) is a protocol that enables LLM agents to self-evolve through natural-language memory rather than weight updates, addressing the trade-off between task specialization and compute cost of supervised fine-tuning. The system converts experience into atomic knowledge artifacts—Rules, Examples, and Mixed types—and uses Population Broadcast to distribute Top-N memory instances across epochs, avoiding the local optimum traps of single-agent learning. Experiments on Gemini-2.5-Flash-Lite and Llama-4-Maverick show strong model-agnostic gains: average return improves from -85.2 (Zero-shot) to +12.6, critical failure rate drops from 28% to 1.1%, and convergence is 2.4x faster, with returns up to 7.7x versus baselines on the CAGE-2 cyber-defense environment. The authors note limitations including potential semantic saturation of the memory store and unverified applicability to fuzzy creative or emotional tasks. Source: ArXiv 2605.16233.

In current LLM engineering practice, there is a persistent logical gap: how to balance task specialization with computational cost. The traditional path relies on supervised fine-tuning (SFT), but at the cost of extremely high compute consumption and potential catastrophic forgetting.

The FORGE Protocol (Failure-Optimized Reflective Graduation and Evolution) offers a highly compact alternative framework whose essence is fully decoupling model intelligence from memory logic.

1. Mechanism Chain: From Reflection to Broadcast

FORGE's logical architecture can be abstracted as:

E_{total} = \text{Reflexion}_{inner} + \text{Broadcast}_{outer} - \text{Cost}_{grad}

> Where: $E_{total}$ is total system efficiency, $\text{Reflexion}_{inner}$ is individual reflection gain, $\text{Broadcast}_{outer}$ is population broadcast gain, and $\text{Cost}_{grad}$ is the compute reduction from the graduation mechanism.

1. Atomization of Knowledge Artifacts: The system does not store raw logs; instead, reflection converts them into three types of artifacts:

Rules: abstracted heuristic logic.
Examples: few-shot demonstrations for specific tasks.
Mixed: synergistic combinations of both.

2. Population Broadcast: This is the system's core accelerator. By distributing Top-N memory instances between epochs, the system effectively avoids the local optimum trap that single agents tend to fall into.

> Local Optimum: An agent finds a "good enough" strategy during learning and sticks to it, missing the chance to discover the optimal solution.

2. Data Density Analysis: The Logic Behind the 7.7x Return

Experimental data on models such as Gemini-2.5-Flash-Lite and Llama-4-Maverick shows FORGE is strongly model-agnostic.

| Metric | Zero-shot (baseline) | Reflexion (single-stream) | FORGE (population) | | :--- | :--- | :--- | :--- | | Average return | -85.2 | -22.4 | +12.6 | | Critical failure rate | 28% | 12% | 1.1% | | Convergence speed | N/A | 1.0x | 2.4x |

3. Conclusions and System Limitations

FORGE demonstrates that: the density ceiling of natural-language memory is far higher than we currently assume.

However, the system still faces the risk of "semantic saturation"—when the error notebook grows thick enough, do additional lessons still yield marginal utility? Moreover, current experiments focus mainly on strongly logical game environments like CAGE-2. Whether this failure-optimized logic remains applicable in fuzzier creative or emotional tasks requires further validation.

---

📚 Paper Details