What access constraints allow description-based adaptation but block conventional techniques?
This explores the specific gap where you can't reach the target — no labeled examples, no access to the model's weights, no sample collection — yet a plain-language description of the new domain is still enough to adapt the system.
This explores the specific gap where you can't reach the target — no labeled examples, no access to the model's weights, no sample collection — yet a plain-language description of the new domain is still enough to adapt the system. The constraint that makes description-based methods shine is *zero-target-access*: you know what the new domain looks like, but you can't get documents or queries from it to fine-tune on the normal way. The clearest case is retrieval adaptation, where a brief textual description of a domain is used to synthesize training data, beating baselines precisely in the scenarios where conventional fine-tuning has nothing to chew on Can you adapt retrieval models without accessing target data?. Description substitutes for the missing collection.
There's a useful map for why this works: specialization techniques live on a ladder of access — black-box (you only get to send prompts), grey-box (you see some internals like token probabilities), and white-box (you can change weights) Does model access level determine which specialization techniques work?. Conventional adaptation lives at the white-box end and gets blocked the moment you lose weight access or training data. A domain description is a black-box move — it activates and steers knowledge the model already has rather than injecting new parameters. That's also the ceiling: black-box methods can only surface existing capability, so when the new knowledge genuinely isn't in the model, description alone won't conjure it.
The same access logic shows up in agents that adapt without ever touching their weights. Reflexion stores verbal self-critiques in episodic memory and improves across attempts on nothing but a binary success/failure signal Can agents learn from failure without updating their weights?, and AgentFly formalizes the whole learning loop as memory operations, reaching strong benchmark scores with the underlying model frozen Can agents learn continuously from experience without updating weights?. When you can't do gradient updates, text becomes the adaptation channel. Fast-Slow Training makes this explicit by deliberately routing task-specific lessons into optimized prompts instead of weights — and finds that doing so dodges the catastrophic forgetting that conventional fine-tuning incurs Can splitting adaptation into two channels reduce forgetting?.
The quiet catch worth knowing: description-based and context-based adaptation assume the model will actually *use* what you tell it. But models often ignore in-context information when their training priors are strong, generating outputs that contradict the very description you supplied Why do language models ignore information in their context?. So the access constraint cuts both ways — text gets you in the door when data and weights are off-limits, but a textual instruction only steers a model that's willing to be steered. Description-based adaptation is best understood not as a weaker substitute for fine-tuning, but as the technique matched to a different constraint: you have words, not data, and you can prompt, not retrain.
Sources 6 notes
Research demonstrates that a brief textual domain description suffices to generate synthetic training data for retrieval fine-tuning, outperforming baselines in zero-target-access scenarios and enabling adaptation where conventional methods are blocked.
Three tiers of access—black-box, grey-box, and white-box—create a hierarchy of specialization power. Black-box techniques can only activate existing knowledge; white-box methods can inject new knowledge but risk over-specialization.
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.
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.
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.
Research demonstrates that LMs generate outputs inconsistent with their context because parametric knowledge from training dominates over in-context information. Textual prompting alone cannot override strong priors; causal intervention in representations is required.