INQUIRING LINE

A skill library is only as useful as how well you sliced the task before you went looking.

What makes skills suitable for retrieval and chaining in repositories?

This explores what actually determines whether a stored 'skill' (a reusable procedure, tool, or knowledge unit) can be found again and stitched into a working chain — not just what makes a repository big, but what makes it usable.


This explores what actually determines whether a stored skill can be found again and chained into a working sequence — and the corpus keeps pointing at the same culprit: the bottleneck is rarely the storage, it's how the task gets broken up before retrieval even begins. The single sharpest finding is that decomposition granularity governs everything downstream: standard LLM decomposition hits only 34% step-level recall, and simply correcting the step count recovers most of the gains, which shifts the real problem to representation-level reranking rather than matching vocabulary What blocks skill retrieval in task decomposition?. In other words, a skill is retrievable when the task has been carved at the joints that match how skills were stored — get the granularity wrong and no amount of embedding tuning saves you.

That connects to a deeper limit the corpus is blunt about: embedding-based retrieval measures association, not usefulness, and there are mathematical ceilings on how many documents a fixed embedding dimension can even represent Where do retrieval systems fail and why?. So 'suitable for retrieval' can't mean 'sits near the query in vector space.' Two approaches close this relevance-vs-usefulness gap. One trains the retriever on whether a skill actually improved the final answer, propagating generation loss back through continuous document representations Can retrieval learn what actually helps answer questions?. The other flips who does the choosing entirely: instead of a retriever guessing, the model itself emits structured tool requests iteratively and refines them as reasoning unfolds, bypassing the colloquial-to-formal vocabulary mismatch that trips up single-round semantic matching Can models decide better than retrievers which tools to use?.

Chaining is a separate skill from retrieval, and the corpus treats it as a form of test-time scaling. Retrieval can be extended into multi-step chains the same way chain-of-thought extends reasoning — generate intermediate retrieval steps via rejection sampling, then dial compute up or down with chain length and search width Can retrieval be extended into multi-step chains like reasoning?. But what makes those chains reliable is how they're trained: rewarding the final answer alone is weak; supervising each intermediate step, and contrasting good chains against bad ones, substantially outperforms outcome-only reward Does supervising retrieval steps outperform final answer rewards?. A skill is 'chainable' when the system has learned which partial sequences lead somewhere, not just which endpoints scored.

The most direct answer to the repository question, though, comes from the idea that a skill library shouldn't be a passive dumping ground. When a trainable curator is decoupled from the frozen executor and learns to evolve the repository, the stored skills drift away from generic verbose additions toward actionable execution logic and cross-task meta-strategies — and that curator generalizes across different executor backbones Can a separate trained curator improve skill libraries better than frozen agents?. Suitability, then, is something you cultivate, not something a skill is born with. A related mechanism gates what's even allowed in: only write back generated material after it passes entailment, attribution, and novelty checks, so the repository grows without polluting future retrievals Can RAG systems safely learn from their own generated answers?.

The thing you might not have expected to want: the payoff isn't a better search algorithm, it's matching the *shape* of stored knowledge to the query's demands. Routing a query to the right structure — table, graph, algorithm, catalogue, or plain chunk — beats uniform retrieval, grounded in the cognitive-science idea that the right representation lowers cognitive load Can routing queries to task-matched structures improve RAG reasoning?. A skill is suitable for retrieval and chaining not when it's well-written in isolation, but when its granularity, representation, and structure were chosen for how it will later be found and combined.


Sources 9 notes

What blocks skill retrieval in task decomposition?

Standard LLM decomposition reaches only 34% step-level recall, gating retrieval success. Correcting step count recovers 75% of gains in iterative methods, shifting the bottleneck to representation-level reranking rather than vocabulary alignment.

Where do retrieval systems fail and why?

RAG systems fail at three structural levels: adaptive triggering (fixed intervals waste context), semantic-task mismatch (embeddings measure association, not relevance), and mathematical limits (embedding dimension constrains representable document sets). These require fundamentally different retrieval approaches, not tuning.

Can retrieval learn what actually helps answer questions?

CLaRa propagates generator loss back through continuous document representations, allowing retrievers to optimize for documents that actually improve answers rather than surface similarity. The gap between relevance and usefulness closes when retrieval receives direct feedback from generation success.

Can models decide better than retrievers which tools to use?

MCP-Zero shows that letting models emit structured tool requests iteratively across conversations outperforms single-round semantic matching. The model can refine requirements progressively across domains as reasoning unfolds, bypassing colloquial-to-formal vocabulary mismatch.

Can retrieval be extended into multi-step chains like reasoning?

CoRAG extends chain-of-thought training to retrieval by using rejection sampling to generate intermediate retrieval chains. Test-time compute can scale through chain length and count, creating a compute dial—greedy decoding for speed or tree search for accuracy—just like reasoning-token scaling.

Show all 9 sources
Does supervising retrieval steps outperform final answer rewards?

Fine-grained feedback on intermediate retrieval steps significantly boosts agentic RAG performance compared to final-answer-only rewards. DPO trained with both positive and negative step feedback outperforms PPO and single-direction training by directly contrasting good and bad retrieval chains.

Can a separate trained curator improve skill libraries better than frozen agents?

SkillOS shows that separating a trainable curator from a frozen executor, grouped by task streams, causes skill repositories to shift from generic verbose additions toward actionable execution logic and cross-task meta-strategies. The trained curator generalizes across different executor backbones and domains.

Can RAG systems safely learn from their own generated answers?

Systems can add generated answers to their retrieval corpus when outputs pass entailment verification, source attribution checks, and novelty detection. This prevents hallucinations from polluting future retrievals while allowing genuine knowledge accumulation.

Can routing queries to task-matched structures improve RAG reasoning?

StructRAG demonstrates that selecting knowledge structure type based on query demands—via DPO-trained router choosing among tables, graphs, algorithms, catalogues, and chunks—improves knowledge-intensive reasoning over standard retrieval. The approach grounds this in cognitive load and cognitive fit theory from cognitive science.

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 investigating a still-open question: what actually makes a stored skill retrievable again and chainable into a working sequence inside a repository? Treat the findings below as dated, perishable claims to re-test, not current truth.

What a curated library found — and when (dated claims, not current truth; findings span ~2023–2026):
- Decomposition granularity, not storage, is the bottleneck: standard LLM decomposition hits only 34% step-level recall, and simply correcting the step count recovers most gains — shifting the real problem to representation-level reranking (~2024).
- Embedding retrieval measures association, not usefulness; a fixed embedding dimension has mathematical ceilings on how many documents it can represent (~2024).
- Retrieval extends into multi-step chains like chain-of-thought — generate intermediate steps via rejection sampling, scale compute with chain length/search width (~2025).
- Process-level supervision (rewarding each intermediate step, contrasting good vs bad chains) substantially outperforms outcome-only reward (~2025).
- A trainable curator decoupled from a frozen executor evolves repositories toward actionable execution logic and cross-task meta-strategies, generalizing across backbones (~2026).

Anchor papers (verify; mind their dates): Divide-or-Conquer? (arXiv:2402.15000, 2024); StructRAG (arXiv:2410.08815, 2024); Chain-of-Retrieval Augmented Generation (arXiv:2501.14342, 2025); SkillOS: Learning Skill Curation for Self-Evolving Agents (arXiv:2605.06614, 2026).

Your task:
(1) RE-TEST EACH CONSTRAINT. For every finding, judge whether newer models, training, tooling, orchestration (memory, caching, multi-agent), or evaluation has RELAXED or OVERTURNED it. Separate the durable question from the perishable limit; cite what resolved it, and say plainly where a constraint still holds.
(2) Surface the strongest CONTRADICTING or SUPERSEDING work from the last ~6 months.
(3) Propose 2 research questions that ASSUME the regime may have moved — reconciling granularity, representation, and curation threads.
Cite arXiv IDs; flag anything you cannot ground in a real paper.