Realize What Matters: Principled Context Representation for Large-Scale Reasoning
Michael Theologitis, Dean Light, Shuyue Stella Li, Benjamin Newman, Yulia Tsvetkov, Dan Suciu
cs.CL
Quick verdict
R3CON: principled corpus-representation harness; +19.9pp (CorpusQA) and +8.4pp (Loong) over 9 baselines; solid ablations, baseline parity needs scrutiny.
Summary
The paper proposes R3CON, a harness for constructing task-specific representations of very large document corpora, grounded in the cognitive theory of relevance realization (Vervaeke et al., 2012). From the theory it derives three design principles: extraction of relevant context should be task-specific and enactive (iteratively reinterpreted as corpus-wide understanding evolves); structuring should be flexible (schema chosen dynamically, e.g., tables, hierarchies, graphs); and relevance and structure should jointly form the representation. R3CON operationalizes these via N rounds of task-focused summarization producing per-document relevance snippets, dynamic schema proposal and per-document parsing into Pydantic structures, and a Python-REPL coding agent for answer synthesis. On two multi-document QA benchmarks (CorpusQA 1M-token split; Loong 50k-250k), R3CON with Qwen3.5-35B-A3B reaches 70.64% and 37.43% overall accuracy, exceeding the strongest of nine baselines by 19.88 and 8.42 points, with standard errors reported. Ablations attribute the largest share to relevance extraction. R3CON with 4B/9B models outperforms all 35B baselines, and beats Claude Code with Sonnet-5 on CorpusQA (71% vs 70%) at roughly 3.7x lower cost. Limitations include per-query recomputation cost, only two QA benchmarks, rescaled baseline hyperparameters, a backbone mismatch for MemAgent, and a qualitative, judgment-based mapping of baselines onto the principles.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Design principles drawn from relevance realization — task-specific and enactive relevance extraction, flexible structuring — explain the successes and failures of existing context-representation methods (Tab. 4, §6.5).
- R3CON, instantiating these principles, outperforms nine SOTA baselines by +19.88 pp on CorpusQA and +8.42 pp on Loong with the same 35B backbone (Tab. 1).
- R3CON reduces reliance on model scale: 4B and 9B variants outperform all evaluated 35B baselines (Tab. 8, Fig. 6).
- On CorpusQA, R3CON with Qwen3.5-35B-A3B beats Claude Code with Claude-Sonnet-5 (71% vs 70%) at 3.7x lower cost (Fig. 1).
- Ablations attribute the larger share of gains to relevance extraction (-13.47/-18.1 pp without it) over structuring (-4.46/-7.42 pp) (Tab. 3).
Strengths
Large, variance-supported gains
Tab. 1 / §D.4: 70.64±2.52 vs 50.76±2.76 (CorpusQA) and 37.43±1.35 vs 29.01±1.28 (Loong); the gaps are roughly 7-10x the SEs and hold in all six domains (Tab. 1) and all four reasoning facets (Tab. 2).
Ablations match the mechanism story
Tab. 3: removing relevance extraction costs -13.47/-18.1 pp and drops R3CON to StructRAG-level performance (52.58% vs StructRAG 50.76% on CorpusQA); removing structuring costs -4.46/-7.42 pp — consistent with the §6.5 failure traces (Figs. 14-17).
Practical scale and cost results
Unusually candid reporting
Reproducibility artifacts
Weaknesses
Baseline field comparability
Post-hoc principle taxonomy
Narrow evaluation scope
Under-powered Claude Code comparison
Per-query recomputation cost
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper tackles question answering and multi-hop reasoning over document corpora far beyond LLM context limits, framed as a *representation-construction* problem. Its central thesis: the ad hoc design of representations (graphs, memories, retrieval pools) explains both the successes and failures of existing systems, and the cognitive theory of relevance realization (Vervaeke et al., 2012) yields three actionable principles — relevance extraction should be task-specific and enactive (reinterpreted iteratively as corpus-wide understanding evolves); structuring should be flexible in both semantic organization and data structure; relevance and structure should jointly form the navigable representation. The authors instantiate these in R3CON and claim state-of-the-art results: +19.88 pp over the strongest of nine baselines on CorpusQA (70.64% vs 50.76%, Tab. 1) and +8.42 pp on Loong (37.43% vs 29.01%), superiority of 4B/9B variants over all 35B baselines (Tab. 8), and a cost-accuracy win over Claude Code with Claude-Sonnet-5 on CorpusQA (71% vs 70% at $439 vs $1,625, Fig. 1).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is R3CON?
R3CON is a research harness that answers questions over very large document collections by first building a task-specific representation of them: it iteratively summarizes each document for relevance to the task, proposes a dynamic data schema (tables, hierarchies, graphs), parses documents into it, and then lets a coding agent reason over the result in a Python sandbox. It was introduced by University of Washington researchers in a 2026 arXiv preprint (2609.27173).
How much better is R3CON than existing methods?
On the two benchmarks tested, R3CON reached 70.64% accuracy on CorpusQA (best baseline: 50.76%) and 37.43% on Loong (best baseline: 29.01%) — improvements of 19.88 and 8.42 percentage points over the strongest of nine baselines, using the same Qwen3.5-35B-A3B model. Smaller 4B/9B versions still beat all 35B baselines.
What is relevance realization?
Relevance realization is a cognitive-science theory (Vervaeke, Lillicrap and Richards, 2012) describing how humans, despite limited working memory, zero in on the information and structure relevant to their current goal. The paper borrows it to argue that AI systems should extract context in a task-specific way and repeatedly reinterpret documents as their understanding of the whole corpus evolves.
Is the R3CON code available?
Yes. The paper links a method repository (github.com/michaeltheologitis/r3con) and an evaluation repository (github.com/michaeltheologitis/r3con-evaluation), and publishes all prompts and baseline implementation details in its appendices.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Section 5 (experimental setup) and parts of the appendices were truncated in extraction; the default N and per-domain sample sizes are not visible in the provided text.
- Figures are available only as embedded text fragments; Pareto/cost trends were read from captions and cited numbers, not from the plots.
- §D.2.5 (effect of rounds N) is referenced but its content is missing from the provided text.
- References are partially truncated; some 2026 citations could not be verified from the text alone.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.