💬 NLP

DIVERGE: Diversity-Enhanced RAG for Open-Ended Information Seeking

The paper introduces DIVERGE, a novel agentic RAG framework that employs reflection-guided generation and memory-augmented iterative refinement to overcome the diversity limitations of standard systems, thereby achieving a superior diversity-quality trade-off for open-ended information seeking.

Original authors: Tianyi Hu, Niket Tandon, Akhil Arora

Published 2026-02-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tianyi Hu, Niket Tandon, Akhil Arora

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

The Problem: The "Echo Chamber" Effect

Imagine you ask a very smart, well-read librarian (an AI) for advice on how to improve your coding skills. You expect a list of different strategies: maybe one person suggests building small projects, another suggests teaching others, and a third suggests studying specific algorithms.

However, current AI systems often act like a broken record. Even if you give the librarian a huge stack of diverse books (retrieved information) to look at, the AI tends to ignore the variety and just give you the same answer it would have given without the books. It collapses all the different possibilities into one "safe," standard response.

The paper calls this "Knowledge Collapse." It happens because the AI is trained to be confident and consistent, so it ignores the "messy" but interesting alternative viewpoints found in the search results.

The Solution: DIVERGE (The "Debate Club" Approach)

The authors created a new system called DIVERGE. Instead of asking the AI to just "give an answer," DIVERGE turns the process into a structured Debate Club.

Here is how it works, step-by-step:

The First Round (The Baseline): The AI gives a standard answer based on what it finds.
The Reflection (The "Wait, what about...?" moment): The system looks at that first answer and asks, "What viewpoints are we missing?" It explicitly identifies a new angle that hasn't been covered yet (e.g., "We talked about technical skills, but what about soft skills?").
The Targeted Search (The "Specific Mission"): Instead of searching for "coding tips" again, the system sends a new search query specifically designed to find evidence for that missing angle.
The Refinement (The "Editor"): The AI generates a new answer based only on that specific new angle, ensuring it doesn't just repeat the first answer.
The Loop: It repeats this process, constantly checking its "memory" to make sure it isn't repeating itself, until it has a full set of diverse, high-quality answers.

The Analogy:
Think of standard AI as a chef who only knows how to make one type of soup, even if you give them a basket of 50 different vegetables. They just make the same soup again.
DIVERGE is like a chef who looks at the basket, says, "Okay, I made a tomato soup. Now, let's look at the mushrooms and make a mushroom soup," then looks at the peppers and makes a pepper soup. It forces the kitchen to use the whole basket, not just the first ingredient.

How They Measured Success

The paper introduces a new way to grade the AI. Usually, we just ask, "Is the answer right?"
DIVERGE asks two questions:

Is it right? (Quality)
Is it different from the others? (Diversity)

They created a "Harmonic Score" (like a GPA) that balances these two. You don't want 10 answers that are all different but wrong, and you don't want 10 answers that are all perfect but identical. You want 10 answers that are all perfect and different.

The Results

When they tested DIVERGE on real-world questions (like "How can I improve my career?" or "What are the best travel destinations?"):

Standard AI + Search: Still gave repetitive answers.
AI trying to be diverse (without DIVERGE): Gave very different answers, but they were often low quality or made up facts.
DIVERGE: Achieved the "Goldilocks" zone. It produced answers that were 2.5 times more diverse in their overall meaning and 1.6 times more diverse in their specific viewpoints, while keeping the quality of the answers just as high as the standard AI.

The Bottom Line

The paper claims that simply giving an AI more information isn't enough to make it think creatively. You have to build a system that explicitly forces the AI to look for different perspectives and checks its own work to ensure it isn't just repeating itself. DIVERGE does exactly that, allowing users to see a wider range of valid possibilities without sacrificing accuracy.

Technical Summary: DIVERGE – Diversity-Enhanced Retrieval-Augmented Generation for Open-Ended Information Seeking

1. Problem Statement

Current Retrieval-Augmented Generation (RAG) systems are predominantly designed under the assumption that every query has a single, factually correct answer. This paradigm leads to knowledge collapse in open-ended information-seeking scenarios (e.g., brainstorming, decision support, or exploring alternative perspectives), where multiple plausible answers exist.

The paper identifies three specific challenges preventing standard RAG systems from generating diverse outputs:

C1: Single-Answer Bias: RAG paradigms encourage overconfident generation, causing models to prioritize a narrow, high-confidence subset of retrieved contexts while ignoring alternative but plausible information.
C2: Missing Diversity Preservation: Existing mechanisms lack long-horizon strategies to preserve diversity across multiple generations. Consequently, even with diverse retrieved contexts, the model produces highly similar, homogenized outputs due to the inherent tendencies of Large Language Models (LLMs).
C3: Limited Practical Applicability: Many state-of-the-art diversity-enhancing techniques rely on access to token-level logits or specific decoding hyperparameters (e.g., temperature, top-p), which are unavailable or restricted in closed-source frontier LLMs (e.g., GPT-5, o3).

Furthermore, the paper notes that simply increasing retrieval diversity (e.g., via re-ranking) does not translate to diverse generations, and existing prompt-based methods often achieve diversity at the cost of significant quality degradation.

2. Methodology: The DIVERGE Framework

The authors propose DIVERGE (Diversity-Enhanced Retrieval-Augmented Generation), a plug-and-play agentic RAG framework designed to explicitly model diverse viewpoints while maintaining answer quality. DIVERGE operates as an iterative process without requiring access to token logits, making it compatible with arbitrary LLM backbones, including closed-source models.

The framework consists of three core components:

A. Reflection-Guided Viewpoint Generation

To mitigate the single-answer bias (C1), DIVERGE treats latent model features as "viewpoints."

Initialization: The system generates an initial RAG response and summarizes it to extract existing viewpoints.
Reflection: In subsequent iterations, the LLM reflects on the set of previously explored viewpoints.
New Viewpoint Identification: The model identifies a new, insufficiently covered viewpoint that offers a distinct perspective on the original query, ensuring the generation process does not collapse into a single dominant stance.

B. Viewpoint-Aware Diversity Retrieval

To address the lack of factual grounding for hypothetical viewpoints (C2), DIVERGE integrates a retrieval mechanism that queries the open web for evidence specific to the identified viewpoint.

Query Generation: The LLM generates a query conditioned on the new viewpoint.
Diversity-Aware Re-ranking: The retriever employs an extended Maximal Marginal Relevance (MMR) formulation. This scores documents based on:
1. Relevance to the current viewpoint-conditioned query.
2. Diversity relative to documents retrieved in previous iterations (memory).
3. Diversity relative to documents selected within the current iteration.
  This ensures the retrieved evidence supports the specific viewpoint while avoiding redundancy with previously explored information.

C. Viewpoint-Conditioned Generation and Refinement

To bridge the gap between a specific viewpoint and the user's original intent, DIVERGE employs a two-step generation process:

Conditioned Generation: The LLM generates a response explicitly conditioned on both the original query and the targeted viewpoint.
Refinement: The generated response is refined to ensure it remains logically coherent, factually accurate, and directly addresses the original user query, preventing excessive deviation while maintaining the diversity introduced by the viewpoint.

3. Evaluation Metrics

Recognizing that standard metrics relying on ground-truth answers are unsuitable for open-ended tasks, the authors introduce a novel evaluation suite focusing on the diversity–quality trade-off:

Semantic Diversity ( $D_{sem}$ ): Measures the variation in meaning across generated responses using pairwise cosine distance between answer embeddings.
Viewpoint Diversity ( $D_{view}$ ): Decomposes responses into atomic claims (viewpoints) and measures the fraction of mutually non-overlapping claims across the set of responses. This captures intra-response diversity that semantic metrics might miss.
Quality Score: Evaluated using an LLM-as-a-Judge paradigm, assessing factual accuracy, evidence support, internal consistency, and question relevance.
Unified Diversity–Quality Harmonic Score ( $U_{Q}$ ): A harmonic mean of normalized diversity and quality scores, providing a single metric to compare the trade-off performance of different methods.

4. Experimental Results

The framework was evaluated on Infinity-Chat, a real-world benchmark of open-ended queries with no ground truth. Experiments utilized GPT-5-mini and GPT-5.1 models.

Performance: DIVERGE achieved the highest Unified Score across all methods and model configurations.
Diversity Gains: Compared to independent sampling (direct prompting), DIVERGE improved semantic diversity by ~2.5× and viewpoint diversity by ~1.6×.
Quality Preservation: Unlike prompt-based baselines (e.g., Verbalized Sampling) which suffered substantial quality degradation, DIVERGE maintained comparable answer quality, with only a marginal decrease (~0.04) in the quality score.
Baseline Comparison: Standard RAG and simple diversity-enhancing strategies (e.g., context shuffling, multi-query) failed to increase diversity significantly and often resulted in lower unified scores.

5. Significance and Claims

The paper claims that DIVERGE addresses a systematic limitation in current LLM-based systems: the failure to leverage diverse retrieved contexts for open-ended information seeking.

Mitigating Homogenization: By explicitly modeling diversity through reflection and iterative refinement, DIVERGE demonstrates that it is possible to overcome the homogenization tendencies of frontier LLMs without retraining or accessing internal model states (logits).
Practical Applicability: The framework is designed as a "plug-and-play" solution compatible with closed-source models, making it immediately applicable to real-world deployment scenarios where decoding controls are restricted.
Ethical and Societal Impact: The authors argue that promoting diversity in AI responses helps mitigate the risk of homogenized outputs that over-represent dominant social or cultural viewpoints, thereby supporting fairer and more inclusive information access and fostering human creativity.

The work concludes that explicitly modeling diversity is a viable path to balancing the often-conflicting objectives of response variety and factual reliability in open-ended RAG systems.