SkillForge, a reinforcement-learning framework for language-model agents, was reported at higher values than cited comparison methods across ALFWorld, WebShop and AppWorld. In a comparison without cold-start initialization, its reported average was 6.3% higher than SkillRL’s. The supplied text does not give task counts, episode counts or counts of independent runs. The figures are benchmark comparisons; they do not by themselves establish that SkillForge caused the differences.
The design makes skills visible
The system records each skill invocation as a discrete event in the agent’s trajectory, making skill use observable. Its listed components include an explicit <skill_call> invocation, multi-pathway skill induction, deduplication, per-skill effectiveness tracking and an LLM reflexion step that keeps or revises skills. Training used GRPO, with a learning rate of 1e−6, group size of eight, KL coefficient of 1e−3 and rollout temperature of 0.9. A Qwen3-Max teacher updated the skill bank every five training steps.
Benchmark comparisons
Using Qwen2.5-7B-Instruct, reported SkillForge values were 93.6 on ALFWorld, 89.8 for WebShop score and 83.0 for WebShop success. On AppWorld, Task Goal Completion, or TGC, was 23.8 and Scenario Goal Completion, or SGC, was 14.3. Against cited GRPO comparisons, the reported differences were +16.0 on ALFWorld, +10.5 on WebShop score and +16.9 on WebShop success. AppWorld TGC and SGC were listed as 17.9 and 3.57 for GRPO, compared with 23.8 and 14.3 for SkillForge.
An additional comparison used the same Qwen2.5-7B-Instruct backbone for SkillForge and SkillRL. SkillForge’s reported differences were +3.7 on ALFWorld and +10.3 on WebShop success. On AppWorld, its TGC and SGC were 23.8 and 14.3, compared with 19.0 and 5.36 for SkillRL.
At other model scales, Qwen3-4B reached 87.9 on ALFWorld and 84.0 on WebShop success, while Qwen3-30B-A3B reached 94.3 on ALFWorld and 59.5 on AppWorld TGC. These reported figures were not accompanied by an uncertainty interval or inferential test.
What happened as the system trained
In component-removal tests using Qwen3-4B, the full system was reported at 87.9 on ALFWorld and 44.6 on AppWorld. The variant without explicit calling was reported at 77.9 and 33.3, the variant without the skill bank at 79.3 and 34.5, and the variant without multi-pathway induction at 82.1 and 36.9, with each pair referring first to ALFWorld and then to AppWorld. Variants without deduplication, effectiveness tracking and LLM reflexion were reported at 86.4 and 38.7, 83.6 and 36.3, and 82.1 and 39.3, respectively. Every listed variant had lower reported values than the full system, but these comparisons alone do not show that any one component caused the difference.
During training, the reported number of active skills grew from 44 to 90 on ALFWorld and from 44 to 86 on AppWorld. The authors list possible skill-bank growth and retrieval overhead as limitations, along with extra tokens that may increase prompt length and inference cost in large deployments.
In a cross-scale transfer analysis, later-stage Qwen3-4B skill banks scored 32.9 versus 27.9 on ALFWorld and 31.5 versus 27.4 on AppWorld. At step 80, the AppWorld comparison was 31.5 for the transferred skills versus 30.4 for the larger model’s self-evolved bank. No uncertainty interval or inferential test was reported for these comparisons.
Training curves showed an initial lag for SkillForge versus a version without explicit calling. Later, SkillForge converged faster, ended with higher reported performance and showed a widening gap. The figure was descriptive and did not report uncertainty bands or repeated-run variation.
The report also compared different teacher settings. With Qwen3-4B fixed as the policy, WebShop values were 79.4 with no teacher, 82.0 with a self-teacher and 84.0 with Qwen3-Max. Corresponding ALFWorld values were 79.3, 85.0 and 87.9, while AppWorld values were 34.5, 39.3 and 44.6. Qwen3-Max had the highest reported value in each of the three cited benchmarks, but no uncertainty interval or inferential test was reported.
The comparisons leave important questions open
An efficiency breakdown listed benchmark scores alongside training time. On AppWorld, SkillForge’s score was 44.6 and its training time was 16.5 hours, compared with 34.5 and 16.1 hours for GRPO. On ALFWorld, SkillForge was 87.9 with 5.9 hours, versus 79.3 with 6.9 hours for GRPO. On WebShop, SkillForge was 84.0 with 7.0 hours, versus 79.4 with 7.2 hours for GRPO. Total skill-related ratios were 4.18% on AppWorld, 9.48% on ALFWorld and 9.47% on WebShop.
The evidence is limited to ALFWorld, WebShop and AppWorld. The supplied text does not report task counts, episode counts, independent-run counts, uncertainty intervals or inferential tests. Limitations also include dependence on an external teacher LLM, possible retrieval overhead as the skill bank grows and extra tokens that may increase prompt length and inference cost. The benchmark differences should therefore be read as reported comparisons, not proof that SkillForge itself caused the observed gains.
The supplied document is an arXiv version 1 preprint, identified as arXiv:2608.24747v1 and dated 25 August 2026. No funding statement is reported in the supplied text.
Paper data and sources
Original title: SkillForge: Evolving Verifiable Skills for Reinforcement Learning Agents
Authors: Shidong Yang, Ziyu Ma, Tongwen Huang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text