INQUIRING LINE

Why do skills an AI agent picks up through training stay locked to the exact scenario it trained on?

What stops evolved agent behaviors from generalizing beyond specific tasks?

This explores why behaviors an agent picks up through training or optimization often stay glued to the exact task they were learned on — and what the corpus says the culprits are.


This explores why behaviors an agent picks up through training or optimization often stay glued to the exact task they were learned on. The corpus points to a few distinct bottlenecks, and the interesting part is that they're not the same bottleneck wearing different hats — they're different failure modes, and the fix for one won't touch the others.

The first is a data ceiling. When agents learn from static expert demonstrations, their competence is capped by whatever scenarios the dataset's curator happened to imagine — they never touch the environment during training, so they can't learn from their own failures or stretch past what was shown Can agents learn beyond what their training data shows?. Generalization can't exceed a ceiling nobody built above the training set.

The second is subtler and almost the opposite problem: reinforcement learning actively narrows behavior. Instead of a ceiling, you get a funnel — RL pushes policies to converge on a few reward-maximizing strategies, collapsing the behavioral diversity that would let an agent adapt to a new task. This 'entropy collapse' shows up in search agents the same way it does in reasoning, and notably, supervised fine-tuning on diverse demonstrations preserves the exploration breadth that RL squeezes out Does reinforcement learning squeeze exploration diversity in search agents?. So the very process meant to make an agent good at a task can make it worse at everything adjacent.

What's striking is that the corpus's success stories route *around* weight updates entirely. VOYAGER stores executable skills in an indexed library and composes complex skills from simpler ones, learning continuously without the catastrophic forgetting that weight-based methods suffer Can agents learn new skills without forgetting old ones?. Agent Workflow Memory generalizes better precisely because it abstracts *sub-task* routines rather than whole tasks — stripping out example-specific values so the routine transfers, with gains that grow as the train-test gap widens Can agents learn reusable sub-task routines from past experience?. The lesson hiding here: generalization improves when you store behavior at a *finer* granularity than the task, not the task itself. Reflexion and AgentFly make the same bet from the memory side — improvement lives in episodic memory or a memory-augmented decision process, not in the frozen weights Can agents learn from failure without updating their weights? Can agents learn continuously from experience without updating weights?.

A survey framework ties this together: self-improving agents split into a slow parametric loop that rewrites model weights and a fast non-parametric loop that rewrites prompts, memory, and tools — and progress is concentrating in the fast loop because scaffold edits are cheap and reversible while weight updates are neither Do self-improving agents really split into two distinct loops?. But even the fast loop has a ceiling you might not expect: the ability to *produce* useful harness edits is flat across model sizes, yet the ability to *benefit* from them peaks at mid-tier models — weak models can't invoke the improvements, and the strongest ones struggle to follow their own instructions faithfully Do stronger models always evolve harnesses better?. So what stops evolved behavior from generalizing isn't one wall — it's a data ceiling, a diversity funnel, and the choice of *where* you store what was learned, with an added twist that being a stronger model doesn't automatically let you cash in your own upgrades.


Sources 8 notes

Can agents learn beyond what their training data shows?

Agents trained on static expert datasets cannot learn from their own failures or generalize beyond demonstrated scenarios because they never interact with environments during training. Competence is capped by what curators imagined, not by agent capacity.

Does reinforcement learning squeeze exploration diversity in search agents?

RL training compresses behavioral diversity in search agents through the same entropy collapse mechanism documented in reasoning—policies converge on narrow reward-maximizing strategies. SFT on diverse demonstrations preserves exploration breadth, suggesting diversity-preservation techniques are essential for RL search scaling.

Can agents learn new skills without forgetting old ones?

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.

Can agents learn reusable sub-task routines from past experience?

Agent Workflow Memory induces sub-task routines at finer granularity than full tasks, abstracts example-specific values, and compounds them hierarchically. This produces 24.6% relative gain on Mind2Web and 51.1% on WebArena, with larger gains as train-test gaps widen.

Can agents learn from failure without updating their weights?

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.

Show all 8 sources
Can agents learn continuously from experience without updating weights?

AgentFly formalizes agent learning as a Memory-augmented MDP with three memory modules (case, subtask, tool) that enable credit assignment and policy improvement entirely through memory operations. The approach achieved 87.88% on GAIA validation without modifying LLM parameters.

Do self-improving agents really split into two distinct loops?

A survey framework organizes self-improving agents into two update mechanisms: slow parametric loops updating foundation model weights, and fast non-parametric loops updating prompts, memory, and tools. Recent progress concentrates in the fast loop because scaffold updates are cheaper and reversible than weight updates.

Do stronger models always evolve harnesses better?

Model capability to produce useful harness edits stays constant across tiers, but capacity to actually benefit from those edits follows an inverted U-shape, peaking in mid-tier models. Weak models fail to invoke harnesses; strong models struggle with faithful instruction-following.

Papers this line draws on 8

The research behind the notes this line reads — ranked by how closely each paper relates.

Research prompt for your LLMexpand ↓

Copy into ChatGPT or Claude to take this line of inquiry further — it asks the model to find newer work and re-test which earlier constraints still hold.

You are a research analyst. Still-open question: what stops evolved agent behaviors from generalizing beyond the specific tasks they were learned on?

What a curated library found — and when (dated claims, not current truth). These findings span roughly 2024–2026:
- Data ceiling: agents trained on static expert demonstrations are capped at scenarios the curator imagined; they never touch the environment, so they can't learn from their own failures (~2025).
- Diversity funnel: RL drives 'entropy collapse,' converging policies on a few reward-maximizing strategies, while SFT on diverse demonstrations preserves the exploration breadth RL squeezes out (~2026).
- Storage granularity: memory that abstracts reusable SUB-task routines (stripping example-specific values) transfers better than task-level memory, with gains growing as the train-test gap widens (~2024).
- Slow vs. fast loop: self-improvement splits into a slow parametric loop (weights) and a fast non-parametric loop (prompts, memory, tools); progress concentrates in the fast loop (~2025).
- Harness twist: the ability to PRODUCE useful harness edits is flat across model sizes, but the ability to BENEFIT peaks at mid-tier models (~2026).

Anchor papers (verify; mind their dates): Agent Workflow Memory (2024); Self-Evolving AI Agents survey (2025); Agent Learning via Early Experience (2025); Harness Updating Is Not Harness Benefit (2026).

Your task:
(1) Re-test each constraint: for every finding, judge whether newer models, training, tooling, orchestration (memory, multi-agent), or evaluation has RELAXED or OVERTURNED it. Separate the durable question from the perishable limit; say where a constraint still holds.
(2) Surface the sharpest DISAGREEMENT in the last ~6 months — e.g., does RL truly collapse diversity, or do diversity-preserving RL methods refute that?
(3) Propose 2 research questions that assume the regime has moved.

Cite arXiv IDs; flag anything you cannot ground in a real paper.