Preprint compares three ways to stop agents forgetting earlier skills

single source· 1 articles · confidence: medium · first seen 2026-09-20 20:00 UTC

What this means for you

If you are fine-tuning one agent for several jobs in sequence, the useful part is the comparison, not the recipe: three recovery methods are pitted against each other, and the answer to forgetting is replay plus a router over separate small weight sets. Nothing to adopt yet — no code, weights or scores are published.

Researchers posted a preprint describing ACLArena, a testbed for what they call agent continual learning: training one agent through several stages of different skills without losing the earlier ones. It analyses where forgetting happens, at model and token level, then compares three ways to recover lost ability — distilling from multiple teacher models, distilling from the agent's own earlier checkpoints, and merging separately trained models. The proposed recipe replays high-quality earlier trajectories and routes between several LoRA experts (small added weight sets, each tuned with reinforcement learning). Evaluation uses four reasoning and agentic tasks; the abstract reports no scores and names no code release.

Key facts

  • ·ACLArena is posted on arXiv as paper 2609.23989, dated 20 September 2026. source
  • ·The framework compares multi-teacher on-policy distillation, self-distilled fine-tuning and model merging as ways to recover capabilities learned at earlier stages. source
  • ·Its proposed recipe combines offline replay of high-quality trajectories with a routed network of multiple LoRA experts, each specialised via reinforcement learning. source
  • ·Evaluation covers four reasoning and agentic tasks, under both in-domain and out-of-domain settings. source
  • ·The abstract reports no benchmark scores, no evaluation dates or harness details, and no released code or weights. source

What the sources say

  • Hugging Face Daily Papers — Single abstract for a preprint; sets out the forgetting analysis and the replay-plus-routed-experts recipe without numbers.

Sources

The original reporting. Follow these — they did the work.

← the wire