💬 NLP

Is Progressive Disclosure All You Need for Long-Context Agents?

This paper presents the first controlled study of progressive disclosure in long-context agents, revealing that while it offers limited benefits when agents already possess strong retrieval capabilities, it becomes decisive for scaling to multi-book tasks where raw document navigation fails, though adding more than one disclosure level provides no additional advantage.

Original authors: Yifeng He, Yinzhe Zhao, Jicheng Wang, Hao Chen

Published 2026-07-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yifeng He, Yinzhe Zhao, Jicheng Wang, Hao Chen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find a specific fact inside a library that contains millions of books. You have two main ways to tackle this. The first way is to try to memorize the entire library at once, hoping your brain is big enough to hold every single page. The second way is to hire a librarian who doesn't know the books by heart but has a card catalog; you ask the librarian for a keyword, they pull a few relevant pages, and you read those. For a long time, computer scientists have debated which of these two methods works best for artificial intelligence (AI).

However, a third, newer idea has emerged: giving the AI the whole library but teaching it how to use a map. Instead of forcing the AI to read everything or relying on an external librarian, you give the AI a "skill pack." This is like handing the AI a table of contents that describes what each chapter is about. The AI can look at the descriptions, decide which chapter sounds useful, and only then open that specific book to read the details. This method is called "progressive disclosure." It's a bit like a detective who is given a massive case file but is first shown a summary of each folder, allowing them to pick the right folder before diving in. The big question is: does this map actually help the detective solve the case faster and better, or is it just extra paperwork that slows them down?

This paper sets out to test that exact question. The researchers built a digital playground called "LOONGDOC" where they could watch AI agents try to answer questions about long books. They tested three different strategies: letting the AI wander through the raw text with no help, giving it a single-level map (a simple table of contents), and giving it a deep, multi-level map (where every single chapter has its own detailed summary). They ran these tests on different types of AI "brains" and with libraries ranging from one book to twenty books.

The results were surprising and very specific. When the AI was dealing with just one book, the fancy maps didn't really help if the AI was already smart enough to find its way around on its own. For example, one of the AI models they tested (Codex) was so good at scanning text that it could find the right answers just by "grepping" (searching) for keywords, making the map completely unnecessary. In fact, for this smart AI, adding a complex, multi-level map actually made things worse, confusing the system and lowering its accuracy. The researchers found that a simple, single-level map was enough; adding a second layer of depth never helped and sometimes broke the system entirely.

However, the story changes completely when you scale up to a library of twenty books. Here, the "raw" approach where the AI wanders blindly collapses. The AI gets lost, burns through its memory trying to read the wrong books, and fails to answer questions. But the AI with the simple, single-level map? It kept its cool. It used the map to quickly identify the right book, then read only what it needed. In this crowded library scenario, the map didn't just help; it was the difference between success and failure. The researchers conclude that progressive disclosure doesn't make the AI smarter; it just gives it better context. It's a tool that is useless when the AI is already a genius navigator, but absolutely essential when the library gets too big to explore without a guide. Ultimately, the paper suggests that for long documents, a single layer of "progressive disclosure" is the sweet spot, while deeper, more complex hierarchies are usually a waste of time.

Technical Summary: Progressive Disclosure for Long-Context Agents

1. Problem Statement

Long-document question answering (QA) traditionally forces a trade-off between two suboptimal strategies: loading an entire document into the context window (often exceeding effective attention limits) or relying on an external retriever to fetch chunks. Agentic AI proposes a third option: providing the agent with the document path and allowing it to decide what and when to read.

While practitioners have rapidly adopted Agent Skills—a standard for packaging expertise into on-demand folders using progressive disclosure (exposing only what a query needs, from short descriptions to specific passages)—this pattern has been driven by engineering intuition rather than controlled evidence. Specifically, it remains unclear whether progressive disclosure outperforms raw-document navigation for book-length tasks, and how the depth of the disclosure hierarchy (flat vs. recursive) impacts performance. Furthermore, existing literature often conflates two design choices: the depth of the routing hierarchy and the physical location of the per-chunk index (always-loaded vs. load-on-activation).

2. Methodology

The authors introduce LOONGDOC, an interactive agent environment built on the BenchFlow framework, which transforms the static $\infty$ Bench benchmark into a sandboxed file-system task. This allows for the recording of agent trajectories, tool calls, and token usage.

Experimental Design

The study compares three distinct approaches to navigating a book-length document, all sharing an identical fixed set of chunks and descriptions to isolate the "wiring" as the only variable:

Raw-Document Navigation (raw): The agent receives the raw book file in the sandbox with no skill pack. It must navigate, grep, or read the file entirely on its own.
Flat Disclosure (flat): The document is packaged as a single Agent Skill. A root SKILL.md file (always in context) contains a table of contents indexing chunk files. The agent reads the root description, activates the skill, and then loads specific chunk files from a references/ directory on demand. Per-chunk descriptions are loaded only upon activation.
Hierarchical Disclosure (hierarchical): A recursive structure where every chunk is its own skill. A meta-router (always in context) routes to child skills, each of which has its own SKILL.md with a description that is always in context. This increases the "always-loaded" context tax.

Evaluation Setup

Benchmarks: $\infty$ Bench subsets: English Multiple Choice (En.MC), English Open QA (En.QA), and Chinese Open QA (Zh.QA).
Scale: Experiments run on a single book and scaled to a library of $K \in \{5, 10, 20\}$ books.
Models & Harnesses: Three agent harnesses (Codex, Pi, Claude-Code) driving three model families (gpt-5.4-mini, qwen3.6-27b, claude-haiku-4.5).
Baseline: A classical Hybrid RAG (retrieve-and-rerank) pipeline serves as an external reference.

3. Key Contributions

Controlled Comparison: The first systematic study isolating the effects of raw navigation versus flat and hierarchical progressive disclosure on a standardized long-document benchmark.
LOONGDOC Environment: A reproducible testbed that converts static book QA into interactive agent tasks, logging full trajectories to explain how agents fail or succeed, rather than just reporting final accuracy.
Design Isolation: The study disentangles the effects of routing depth and index location, demonstrating that the "book-to-skill" recipe's success depends heavily on the agent's native navigation capabilities and corpus scale.

4. Results

Single-Book Performance

Harness Dependence: Progressive disclosure is not a universal accuracy lever.
- Weak Navigators (Pi, Claude-Code): Flat disclosure matches or exceeds raw navigation. The skill pack helps agents that struggle to locate passages in raw text.
- Strong Navigators (Codex): For agents like Codex (gpt-5.4-mini), which naturally "grep" for entities and read only matched passages, the skill pack adds zero accuracy gain. The agent reconstructs the locate-then-read capability on the fly, making the pre-cut skill redundant.
Depth Penalty: The hierarchical approach (deeper recursion) never outperforms the flat approach on a single book and often degrades accuracy significantly (e.g., dropping En.MC from 0.91 to 0.63 on Pi). The constant context load of always-loaded child descriptions saturates the router's context before it commits to a chunk.

Library-Scale Performance (Scaling to $K=20$ )

Collapse of Raw Navigation: As the corpus grows, raw-document navigation collapses. For strong agents like Codex, accuracy on open QA drops drastically (e.g., En.QA falls from ~0.66 to ~0.26) because the agent cannot reliably locate the correct book among many without burning excessive context.
Flat Disclosure as a Rescue: The flat skill pack degrades much more slowly and pulls ahead of raw navigation at scale. It allows the agent to commit to the target book before traversing the library.
Hierarchical Failure at Scale: The hierarchical pack fails to reproduce the rescue seen in the flat pack. The "always-loaded" descriptions of all chunks across 20 books inflate the context budget, recreating the pressure the method aims to relieve.
Efficiency: The flat pack is not only more accurate but more efficient. At $K=20$ , it achieves nearly double the accuracy of raw navigation at roughly half the token cost.

Baseline Comparison

Progressive disclosure (specifically flat) generally outperforms the classical Hybrid RAG baseline, particularly on open QA tasks, suggesting that in-skill indexing is a more effective routing mechanism than standard retrieve-and-rerank pipelines for structured long documents.

5. Significance and Claims

The paper concludes that progressive disclosure buys context, not intelligence.

Redundancy vs. Decisiveness: For a single book, if an agent is strong enough to navigate raw text effectively, progressive disclosure is redundant. However, once the corpus grows too large for native navigation (library scale), progressive disclosure becomes decisive.
Optimal Structure: One level of disclosure (flat) is sufficient. Adding a second, deeper routing level (hierarchical) never helps and often breaks accuracy by saturating the context window with metadata.
Practical Guidance: For packaging book-length material, the authors recommend a single skill layered with progressive disclosure (a flat index) rather than a deep hierarchy of child skills.

The study emphasizes that the benefits of this pattern are task- and language-specific, showing the sharpest gains on English open QA and negligible or negative effects on Chinese QA (where base model capability sets the ceiling) and multiple-choice tasks (where pre-training memorization confounds the results).