Adaptive Context Selection for Retrieval Augmented Models

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2149 papers190 conceptsfrozen 2026-10-04

In standard Retrieval-Augmented Generation (RAG), the system follows a rigid recipe: take a query, fetch exactly $k$ documents, and stuff them into the prompt. This "fixed-k" approach assumes that every question requires the same amount of evidence. However, a query like "What is the capital of France?" needs far less context than "Compare the economic policies of the 1920s and the 1990s."

A growing body of research suggests that static retrieval constraints are a primary source of both "hallucinations" (when context is insufficient) and "noise distraction" (when context is excessive). The emerging direction is Adaptive Context Selection: mechanisms that dynamically determine the volume, granularity, and necessity of external information based on the specific demands of the task.

The Fixed-k Bottleneck

Most retrieval systems treat documents as independent, static units. As noted in Efficient and Robust Question Answering from Minimal Context over Documents by Sewon Min and colleagues, the vast majority of questions can be answered with just one or two sentences. Forcing a model to read ten full paragraphs introduces irrelevant information that can lead the model astray.

Conversely, for complex tasks, a fixed budget might truncate the very evidence needed for a complete answer. Lang Qin and the authors of Well Begun is Half Done argue that traditional retrieval fails because it doesn't account for "dialogue ambiguity." They propose GATE, a module that adjusts the quantity of retrieved facts based on how certain the system is about the user's intent.

Scaling Volume to Query Complexity

The first facet of this research involves Dynamic Volume Selection—autonomously deciding how many pieces of evidence to retrieve.

In AdaComp, Qianchi Zhang and others use a lightweight predictor to estimate the minimum number of documents required to answer a query. This prevents "information mix-up" while slashing inference costs. Similarly, EviRerank by Minghan Li introduces Adaptive Evidence Budgeting, which uses a relative score threshold. Instead of taking the top 10 results, the system stops including segments when their relevance score drops significantly compared to the top result, effectively filtering out low-utility noise.

This principle extends beyond text. In Risk Controlled Image Retrieval, Chris Xiaoxuan Lu demonstrates that for safety-critical tasks, the system should return a variable-sized set of results to guarantee a "true match" is included, rather than a fixed number of potentially wrong guesses.

The Granularity Trade-off

Beyond how much to retrieve, researchers are asking at what scale we should retrieve. This is the problem of Chunking Granularity.

Small chunks (sentences) offer high precision but lack context; large chunks (paragraphs) offer context but include noise. Zijie Zhong and the team behind Mix-of-Granularity (MoG) address this by using a router to select the ideal chunk size for each query. A technical query might trigger a narrow, precise snippet, while a narrative query triggers a broad, context-heavy segment.

Xuechen Zhang further refines this in SmartChunk, using reinforcement learning to predict the necessary "retrieval window" for a specific request. By dynamically expanding the context window around a retrieved fact only when initial evidence is insufficient, systems like OkraLong capture nuances that fixed-chunk systems miss.

Selective Retrieval and Self-Correction

The most advanced adaptive systems don't just decide how much to retrieve—they decide if they should retrieve at all.

REPOFORMER, developed by Di Wu and colleagues, trains models to self-evaluate their internal knowledge. If a model is confident, it skips retrieval entirely, avoiding the "distraction" of external data and speeding up inference by 70%.

This "think-before-you-act" philosophy is also seen in:

  • In-Context Learning: Fei Zhao's $D^2Controller$ dynamically selects the optimal number of few-shot examples, preventing performance degradation from "example overload."
  • Attribution: Emil Vatai's SAFE framework predicts how many references a sentence needs before searching for them, ensuring citations are precise rather than performative.
  • Schema Linking: Zheng Yuan's KaSLA treats context selection as a knapsack optimization problem, balancing a "redundancy budget" against query complexity to prevent LLMs from being overwhelmed by database schemas.

Why It Works: The Entropy of Information

The intuition behind adaptive selection is that information density is non-uniform. As Zhi Gao notes in MemoryEQA, high-entropy (ambiguous) queries require a wider context window to resolve uncertainty, while low-entropy tasks achieve higher precision with a narrower, high-confidence set. By moving from static "brute-force" reading to dynamic, query-aware selection, these models mimic human-like efficiency—reading only what is necessary to reach a confident conclusion.

Papers in this version

Showing 60 of 149 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 124 papers
  2. Version 1 · 2026-10-04 · 149 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_dynamic_context_windowing_011b1ebb,
  title = {dynamic context windowing},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dynamic-context-windowing/rdr_011b1ebbb2e34277ae4fa57ea9902bd7},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_adaptive_context_selection_for_retrieval_011b1ebb_v1,
  title = {Adaptive Context Selection for Retrieval Augmented Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adaptive-context-selection-for-retrieval-augmented-models/rdr_011b1ebbb2e34277ae4fa57ea9902bd7/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}