Automated literature reviewers now read thousands of papers, but they routinely compare numbers and findings that are not comparable. A sentence like "method A reaches 71% while method B reaches 64%" is only meaningful if both numbers come from the same dataset split, metric variant, and model scale. In a paper, those conditions usually sit somewhere else: a definition in Section 2, a table caption, a footnote, or a hedge three sentences earlier. Current systems pull the claim and lose its conditions. We propose to make recovering those conditions an explicit, cited retrieval step. Each claim gets "closed" by following its dependencies inside its own paper. The plan for the review is then built from comparison questions that can only be answered once the conditions of the claims involved have been matched or shown to differ.
Two existing observations motivate this. Language understanding has long been known to depend on the surrounding text: word prediction in The LAMBADA dataset becomes impossible without broad context. Whether a statement is generic or about a specific setting (Discourse-sensitive Automatic Identification of Generic Expressions), and whether a modifier restricts its noun (Annotating and Predicting Non-Restrictive Noun Phrase Modifications), are also decided by context. Meanwhile, agents answering multi-paper questions over full-text AI papers top out at 44.14% accuracy on AirQA, against 56.63% for graduate-student experts, and they do worst on tables and formulas. That is exactly where a value depends on a definition located elsewhere in the paper. Our bet is that a large share of the multi-hop gap in scientific synthesis is not a failure to find other papers. It is a failure to carry each claim's scope along with it.
Core Claim
A "condition-closed claim" is an extracted assertion paired with a small set of condition slots. The slots cover dataset or population, experimental setting, metric definition, comparison baseline, and hedging or negation scope. Each filled slot carries a pointer to the source span that supports it, found by following discourse dependencies within the same paper. Our claim is that a synthesis system whose hierarchical plan is driven by condition mismatch will produce fewer invalid comparisons and more correct quantitative answers than systems that drive search by outline headers, critic prompts, or knowledge-graph topology. "Driven by condition mismatch" means two things. A comparison question in the plan counts as unresolved while its claims disagree on an unflagged slot. Each unresolved question triggers either an intra-paper hop, to fill a slot, or an inter-paper search, to find a claim under matching conditions. The new element is the coupling: discourse context recovery becomes the trigger for multi-hop search and the unit of planning, rather than a preprocessing nicety.
Relation to Prior Work
Two systems are closest, and they differ from this proposal in opposite ways. A Tale of Two Graphs (DualGraph) separates an outline graph from a knowledge graph of atomic facts. It drives search by graph topology, filling structural holes between clusters and enriching weakly cited relations, and it beats an outline-only variant on DeepResearch Bench II (41.48 vs 39.50). Its atoms carry no scope conditions, so two facts measured under different metrics can sit on adjacent nodes and look like a coherent edge. We keep its central lesson, that outline headers are weak search supervision, and replace topological gaps with condition gaps. Disco-RAG parses rhetorical structure inside and across retrieved chunks and plans generation from it, gaining 12.74 LLM-score points on the Loong benchmark. It works over a small, fixed retrieved set, and its rhetorical relations (elaboration, contrast) are about how text flows, not about whether two findings were measured under the same conditions. It therefore supplies evidence that explicit discourse structure helps, but not the mechanism we need.
AgenticScholar is the closest in spirit on the comparison side. Its operator for building problem-by-method comparison matrices pulls information from specific paper sections, reaching 0.73 R-Precision against 0.61 for the best retriever baseline. However, it produces matrices rather than prose, and it treats extraction as slot-filling from a section, not as following a dependency chain to where a definition actually lives. Several survey writers organize search and writing by outlines or intent categories over mostly abstract-level evidence. These include IRIS, with a critic-guided intent loop scoring 53.46 RACE on DeepResearch Bench; SurveyMaster, with a metadata-ranked core-paper backbone; and CHIME, with LLM-built study hierarchies. IRIS names abstract-only evidence as its main limitation. CHIME's finding constrains our design. LLMs build sensible category trees (99.9% parent-child precision) but assign individual studies to them poorly (61.5% F1). That suggests plans should not rely on the model's unaided placement of evidence, and condition slots give placement an explicit criterion.
On the method side, we borrow question-shaped plans from Explanatory Summarization with Discourse-Driven Planning. That work turns discourse units into "questions under discussion" (the implicit questions a passage answers) and uses them to condition generation, but only within a single paper. We also borrow the interleaving of reasoning and retrieval from IRCoT and ITER-RETGEN. Those systems form each new query from the previous reasoning step. We form it from an unfilled slot or an unmatched comparison, which is a narrower and checkable target. Our verification design also responds to the warning in Wang et al. (2025) and Attribution Techniques for Mitigating Hallucinated Information in RAG Systems: an early bad retrieval cascades into later errors. Because every slot carries a span pointer, an error can be audited where it starts. GLIMPSE shows that consensus-seeking summaries erase divergent findings. Flagged condition differences are one principled way to report divergence rather than average it away.
Research Program
The system has three parts.
The first is a claim closer. Given a full-text paper parsed into text, tables, captions, and formulas, it extracts candidate claims and, for each, issues intra-document queries for each empty slot. Examples are "where is the metric in Table 3 defined?" and "what split does 'test' refer to here?". It follows coreference and section cross-references until each slot is filled with a cited span or marked absent. The closer starts as a prompted LLM. If annotated slot data suffice, we then train a 7B model on its trajectories, following AirQA's finding that a 7B model trained on 4,000 synthetic trajectories matched an untrained 14B model.
The second is a planner. It turns the user's topic into a tree whose internal nodes are sections and whose leaves are comparison questions under discussion, such as "how do learned indexes compare with B-trees on lookup latency at fixed memory?". Each leaf holds the closed claims assigned to it, and assignment is checked against slot agreement. A leaf is open if its claims disagree on a slot without an explicit flag, or if it holds evidence from only one condition. Open leaves issue intra-paper hops or new corpus searches with the required conditions written into the query. Search stops when leaves close or a budget is exhausted.
The third is a writer that realizes each closed leaf. It has three options: state a direct comparison when conditions match, state a flagged comparison naming the difference when they do not, or state that no comparable evidence was found.
The analysis component asks where invalid comparisons come from in current systems. We annotate the comparative sentences in reports generated by the baselines below and classify each failure. A failure can be a slot that was present in the paper but never retrieved, a slot that was retrieved but ignored in writing, or a slot genuinely absent from the source. This tells us whether the bottleneck is closing claims, planning, or writing.
Baselines
The naive reference point is standard one-shot retrieval over full-text chunks with the same backbone. On AirQA, Qwen2.5-72B with standard RAG scored 20.22%, and title-plus-abstract inputs scored below 10% (AirQA). We rerun this rather than relying on the reported figure.
The strongest reported agentic result on AirQA is 44.14%, the best agentic setup including GPT-4o. The same paper reports 35.07% for its agentic hybrid approach on Qwen2.5-72B. We run that hybrid agent from the released AirQA codebase on the same backbone we use. Beating it means a statistically reliable gain on the multi-document, table, and formula categories, not merely on the overall average. Our bar is to exceed 35.07% on Qwen2.5-72B. With a GPT-4o-class backbone, the bar is the 44.14% reported for the best agentic setup.
For long-form synthesis, the strongest competitors are IRIS and DualGraph. IRIS reports 53.46 RACE on the 50 English DeepResearch Bench tasks, and DualGraph reports 53.08 with GPT-5. Since neither was designed for full text, we reimplement both on our full-text corpus with the same backbone, the same retrieval index, and the same per-report token budget. IRIS reports roughly 569K tokens per report, and we cap all systems near that figure. Every system gets an equal number of prompt-tuning iterations on a held-out development split.
On SciSurveyBench, the strongest reported number is SurveyMaster's 4.84/5 fact correctness and 4.49/5 citation precision, up from 3.63 and 3.62 for the strongest automatic baseline (SurveyMaster). Matching fact correctness while sharply reducing invalid comparisons would count as a win there.
Three ablations isolate the key ingredient. The first removes condition slots entirely, keeping the question-shaped plan. The second fills slots but does not let them open or close leaves, so they are only passed to the writer. The third replaces the intra-paper hops with a single whole-paper rewrite, described in the next section. If the second ablation matches the full system, the planning coupling is not doing the work.
Why not just…?
Why not just put the full papers into a long-context model? Reported evidence points against it. DualGraph documents that linear accumulation of context causes evidence buried in the middle to be ignored, and Disco-RAG's long-context prompting baselines trail it by 12.74 points on Loong. Neither tested condition fidelity specifically, though. Our full-text long-context baseline, run at the same token cost, measures whether the model compares like with like when everything is technically in view.
Why not just ask the model to rewrite each claim as a standalone sentence, with the whole paper in context? This is the real threat. A single rewrite pass may recover most conditions in cleanly written text. We test it directly as the third ablation and measure slot accuracy against human annotations, split by whether the dependency crosses sections or lives in a table. If the rewrite matches explicit hops, we adopt it as the claim closer. The proposal's contribution then narrows to the mismatch-driven planner, which the second ablation still tests independently. We expect explicit hops to win mainly on table and formula claims, where AirQA shows the steepest drops, and on auditability, since only hops yield span pointers.
Why not just tell an IRIS-style critic to check whether claims are comparable? A prompted critic can catch obvious mismatches, but it judges from whatever evidence was already extracted. If extraction dropped the metric definition, there is nothing to check. Our IRIS baseline gets exactly this instruction added to its critic prompt. The failure analysis will show whether its residual errors are "slot never retrieved" cases, which a critic cannot fix by construction.
Proposed Experiments
The first experiment is quantitative multi-paper question answering on AirQA, reported by task type and by element category (text, table, image, formula, metadata). The prediction is that gains concentrate in multi-document and table/formula questions. Flat gains across categories would suggest a generic agent improvement rather than better condition closure.
The second is long-form synthesis. We will run on SciSurveyBench for multi-discipline reviews. We also need a CS-focused set: AgenticScholar describes a corpus of 237 papers across 10 topics with gold taxonomies from 7 ACM Computing Surveys, but the file gives no public release identifier for it. We will request it from the authors, and if it is unavailable, we will substitute AirQA's ACL, ICLR, and NeurIPS full-text papers grouped by topic. Results will be reported with the benchmarks' own judges, plus citation accuracy in the style of the FACT metric from DeepResearch Bench, where Perplexity reached 82.6%. DeepResearch Bench itself is a web-scale benchmark, so we use it only as a sanity check that general report quality does not degrade.
The third, and the decisive check, is a new invalid-comparison rate. Domain annotators will label a sample of comparative sentences from every system's reports as valid, validly flagged, or invalid, given the cited source spans. Because no existing benchmark measures this, we will release the annotations. The hypothesis is falsified if the full system's invalid-comparison rate is not clearly below both the IRIS-style critic and the second ablation at matched token budget, or if its gains on AirQA disappear once we control for the number of retrieval calls. A secondary check tests whether CHIME's placement weakness recurs. We will measure leaf-assignment F1 against annotated slot agreement, and expect slot-checked assignment to beat unaided LLM placement by a margin comparable to CHIME's 12.6-point gain from its trained corrector.
References
- The LAMBADA dataset: Word prediction requiring a broad discourse context
- Discourse-sensitive Automatic Identification of Generic Expressions
- Annotating and Predicting Non-Restrictive Noun Phrase Modifications
- AirQA: A Comprehensive QA Dataset for AI Research with Instance-Level Evaluation
- A Tale of Two Graphs: Separating Knowledge Exploration from Outline Structure for Open-Ended Deep Research
- Disco-RAG: Discourse-Aware Retrieval-Augmented Generation
- AgenticScholar: Agentic Data Management with Pipeline Orchestration for Scholarly Corpora
- IRIS: An Agentic Multi-Phase Framework for Automated Scientific Literature Review
- Automating Cognitive Distillation for Expert-Level Scientific Literature Synthesis
- CHIME: LLM-Assisted Hierarchical Organization of Scientific Studies for Literature Review Support
- Explanatory Summarization with Discourse-Driven Planning
- Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions
- Enhancing Retrieval-Augmented Large Language Models with Iterative Retrieval-Generation Synergy
- Attribution Techniques for Mitigating Hallucinated Information in RAG Systems: A Survey
- GLIMPSE: Pragmatically Informative Multi-Document Summarization for Scholarly Reviews
- DeepResearch Bench: A Comprehensive Benchmark for Deep Research Agents