SkillForge: Co-Evolving Skills and Agents via Dynamic Skill Lifecycles
Introduction
This is the final SkillForge policy for ALFWorld: a full fine-tune of
Qwen2.5-7B-Instruct trained with GRPO while its skill library was simultaneously
forged — retired, stabilized, demoted and mutated — under the fitness-driven
lifecycle described in the paper.
Most memory-augmented agents keep the library append-only. A skill that was
correct at step 20 encodes a procedure the policy has outgrown by step 120, and it is
still being retrieved into the context. SkillForge instead scores each skill against
the policy's own rollouts and moves it between four states — trial, active,
stable, retired — so the library and the model co-evolve.
On ALFWorld this reaches 92.4% overall success against 89.9% for SkillRL, the
strongest memory-augmented RL baseline, and 14.8% for the untuned base model.
What is in this repository. The policy weights and tokenizer only. The evolved
skill library is not shipped here: the agent is the policy plus the retrieved
skills injected into its context. Skill contents, fitness trajectories and
retirement events are documented in SkillFurnace (Appendix C of the paper) and in
the paper's case studies.
Results
ALFWorld, per task type and overall (%).
| Method |
Pick |
Look |
Clean |
Heat |
Cool |
Pick2 |
All |
| Vanilla (base model) |
33.4 |
21.6 |
19.3 |
6.9 |
2.8 |
3.2 |
14.8 |
| ReAct |
48.5 |
35.4 |
34.3 |
13.2 |
18.2 |
17.6 |
31.2 |
| Reflexion |
62.0 |
41.6 |
44.9 |
30.9 |
36.3 |
23.8 |
42.7 |
| GRPO (no library) |
90.8 |
66.1 |
89.3 |
74.7 |
72.5 |
64.7 |
77.6 |
| SkillRL |
97.9 |
71.4 |
90.0 |
90.0 |
95.5 |
87.5 |
89.9 |
| SkillForge |
98.2 |
86.6 |
89.8 |
93.7 |
95.6 |
88.4 |
92.4 |
The gain over SkillRL is largest where the task is loosest: +15.2 on Look, +3.7 on
Heat, +0.9 on Pick2. Those are the task types where a stale instruction hurts
most, which is the failure mode the lifecycle exists to remove.
The library this model ended up with
|
Seed library |
Final library |
| Total skills |
44 |
100 |
| General |
12 |
— |
| Task-specific |
21 |
— |
| Common-mistake |
11 |
— |
The run saturates its skill cap (S_max = 100). For scale, the w/o retirement
ablation — the same method with retirement switched off — grows to 132 skills and
drops to 91.4%: a bigger library is not a better one.
How it was trained
- Pre-retirement. The seed library, inherited from the SkillRL release, is scored
under 500 rollout episodes of the base model with skills injected. Anything whose
proto-fitness falls below
delta_pre = 0.3 is retired before training starts.
- Retirement-aware cold start. The base model is fine-tuned with cross-entropy on
the rollouts that survived, with trajectories that leaned on a since-retired skill
filtered out.
- Skill-policy co-evolution. GRPO takes over and, every 10 steps, the forging
cycle reads each skill's runtime fitness and promotes, demotes, retires or mutates
it. Mutation is LLM-guided (Kimi-K2.5 as teacher); at most 5 mutations and 3
retirements per cycle; retrieval is top-6 by task-type match.
| Hyperparameter |
Value |
| Base model |
Qwen2.5-7B-Instruct |
| Optimizer |
GRPO, via verl |
| Learning rate |
1e-6 |
| Batch size / group size |
16 / 8 |
| Clip epsilon / KL beta |
0.2 / 0.001 |
| Training steps |
200 |
| Sampling temperature (train and eval) |
1.0 |
| Hardware |
64 x NVIDIA H200 |
Quick Start
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "YuyaoGe/SkillForge_Alfworld_7B"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype="bfloat16", device_map="auto")
# Skills are retrieved (top-6, by task-type match) and injected into the context.
# The exact prompt template and the action space are in the paper.
messages = [
{"role": "system", "content": "<retrieved skills for this task type>"},
{"role": "user", "content": "<observation>\n\n> "},
]
ids = tok.apply_chat_template(messages, add_generation_prompt=True,
return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=256, do_sample=False)
print(tok.decode(out[0][ids.shape[-1]:], skip_special_tokens=True))
This is a research artifact: a 7B text policy for a text environment. It exposes no
tool interface of its own and will emit ALFWorld-style actions for anything that
resembles an ALFWorld prompt.
Citation
@misc{ge2026skillforgecoevolvingskillsagents,
title={SkillForge: Co-Evolving Skills and Agents via Dynamic Skill Lifecycles},
author={Yuyao Ge and Yiwei Wang and Yuchen He and Baolong Bi and Lingrui Mei and Jiayu Yao and Lizhe Chen and Shenghua Liu},
year={2026},
eprint={2610.09832},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2610.09832},
}
Acknowledgments
Training runs on verl for the GRPO loop, with
seed skill libraries inherited from the SkillRL release.
