When an AI learns one job, does that skill carry over to new tasks, or vanish when it learns the next?
Does AI training preserve learning that transfers to independent subsequent tasks?
This explores whether what a model learns in training carries into tasks it wasn't trained on, and whether it can then learn the next task without losing the last one.
This explores whether what a model learns in training carries into tasks it wasn't trained on, and whether it can then learn the next task without losing the last one. The corpus says yes, with conditions. It depends on what was learned and on how the training was wired.
Generalization is the encouraging part. AIDE2's gains held on four held-out benchmarks, including physics-based weather forecasting, which sat outside the distribution used to select it Do AIDE2's improvements transfer to unseen tasks?. Training models to hand subtasks to subagents also produced a skill that carried over to plain single-agent work. That suggests it taught disciplined decomposition and evidence-grounding, not just how to orchestrate Can delegation teach models to manage context more actively?. Transfer seems to happen when training instills a way of working, not a task-specific trick.
Some of what looks like learning is narrower than it seems. Models instruction-tuned on semantically empty or deliberately wrong instructions did about as well as those trained on correct ones. What carried over was the shape of the output, not understanding of the task Does instruction tuning teach task understanding or output format?. RL post-training has a similar narrowing effect. Within the first epoch it amplifies one format from pretraining and collapses the alternatives, and which format wins depends on model scale Does RL training collapse format diversity in pretrained models?. So training preserves something, but that something can be quite thin. Moving reasoning training into pretraining itself is one attempt to plant it deeper, and it lifted math and science benchmarks Can chain-of-thought reasoning be learned during pretraining itself?.
The sequential half of the question is about forgetting. The notes locate the problem in where the learning is stored. Fast-Slow Training keeps weight updates small and puts task-specific lessons into optimized prompt context. It reached equal performance 1.4–3x faster with much less forgetting, which frames forgetting as misallocation rather than an unavoidable cost Can splitting adaptation into two channels reduce forgetting?. The models also stayed up to 70% closer to their base distribution. That closeness kept them able to learn later tasks, while parameter-only RL stalled when the domain changed Does staying close to the base model preserve learning ability?. Two agent designs go further and keep learning out of the weights entirely. VOYAGER stores executable skills in a library and composes new ones from old ones Can agents learn new skills without forgetting old ones?. Reflexion stores verbal self-diagnoses in episodic memory Can agents learn from failure without updating their weights?.
Order matters too. Training structured tasks first and creative ones later beat joint training by 6.2%, because structured domains push output entropy down and creative ones push it up Does training order reshape how models handle different task types?. The same question has a less reassuring answer for people. In a four-month EEG study of 54 participants, heavier LLM use went with weaker brain connectivity, poorer memory retention, and trouble recalling one's own recent work Does AI assistance weaken our brain's ability to think independently?. Models can be engineered to keep and transfer what they learn, but nothing suggests the human user keeps what the model does for them.
Sources 11 notes
The paper reports that AIDE2's improvements transfer to four held-out benchmarks spanning machine learning, algorithm engineering, and physics-based weather forecasting—the last being outside the selection task distribution. This demonstrates transferable gains beyond overfitting to the selection set.
SearchSwarm shows that training models to delegate subtasks and integrate summarized results beats passive compression, with a 30B model matching much larger ones. Critically, the delegation skill transfers to single-agent tasks, suggesting it teaches disciplined decomposition and evidence grounding, not just orchestration.
Models trained on semantically empty or deliberately incorrect instructions achieve comparable performance to those trained on full correct instructions, achieving 43% vs random baseline 42.6%. The semantic content of instructions appears largely irrelevant; what transfers is knowledge of the output space.
Controlled experiments show RL consistently amplifies one format distribution from pretraining within the first epoch while collapsing alternatives. The winning format depends on model scale, not necessarily performance, and is largely hidden when starting from proprietary pretrained models.
RLP treats CoT as exploratory action during pretraining, using log-likelihood improvement as verifier-free reward. Applied to Qwen3-1.7B and Nemotron-Nano-12B, the method improves math and science benchmarks substantially, suggesting reasoning can be planted earlier in training.
Show all 11 sources
Fast-Slow Training routes task-specific lessons into optimized prompts while keeping parameter updates minimal, reaching equivalent performance 1.4–3x faster with substantially less catastrophic forgetting and plasticity loss, demonstrating that forgetting is a misallocation problem rather than an inherent cost.
FST-trained models stay up to 70% closer to their base distribution than parameter-only RL, and this reduced drift preserves the model's ability to learn subsequent tasks effectively. Parameter-only approaches stall when task domains change, while low KL drift enables sustained adaptation.
VOYAGER demonstrates that storing executable skills in an embedding-indexed library and composing complex skills from simpler ones allows agents to learn continuously while avoiding the forgetting that occurs with weight-update-based methods. Environmental feedback refines skills while an automatic curriculum drives continual exploration.
Reflexion demonstrates that unambiguous environmental feedback (success/failure) enables agents to write useful self-diagnoses and improve across episodes without parameter updates. The binary signal prevents rationalization, and keeping reflections uncompressed preserves their usability.
Omni-Thinker shows structured domains decrease output entropy while creative domains increase it. BWT-guided scheduling—training structured tasks first—yields 6.2% gains over joint training by preventing entropy collapse from damaging open-ended capabilities.
A four-month EEG study of 54 participants found that brain connectivity systematically scaled down with AI reliance—LLM users showed weakest neural engagement, poorest memory retention, and impaired ability to recall their own recent work.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- SkillClaw: Let Skills Evolve Collectively with Agentic Evolver
- Learning, Fast and Slow: Towards LLMs That Adapt Continually
- A Survey on Post-training of Large Language Models
- Eliciting Reasoning in Language Models with Cognitive Tools
- AgentFly: Fine-tuning LLM Agents without Fine-tuning LLMs
- MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation
- Demystifying Agent Skills: Why They Work-Until They Don't
- On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models