Many disease mechanisms involve no change in how much of a protein a cell makes. What changes is where the protein sits. A tumor suppressor can be intact and abundant and still do nothing, because it is held in the cytoplasm and never reaches the DNA it is supposed to regulate. Biologists call this cytoplasmic sequestration. Reconstructing such a mechanism from the literature is a chain of lookups across several papers, for example "what stabilizes the binder, what does the binder hold, where does the held protein end up, which genes change as a result." The decisive evidence for each link is usually a figure: a Western blot that compares the nuclear fraction with the cytoplasmic fraction, or a tissue stain scored separately by compartment. Systems that answer multi-step questions by retrieving documents, and systems that read scientific figures, each handle half of this problem. Neither carries the qualifier that matters most here: which compartment, in which genetic background, under which condition.
We propose that the dominant failure of multi-step literature question answering in this setting is qualifier dropping. The system links two hops through the protein name and loses the compartment and genotype that made the first hop true. It then joins evidence that does not belong together, or reaches the right answer for the wrong reason. The method keeps an explicit record of these qualifiers at every step of the retrieval chain and requires each new hop to match them. It also opens high-resolution views of figures only when a qualifier cannot be settled from text. The project builds a benchmark designed to expose qualifier dropping, and tests whether qualifier-carrying chains fix it more cheaply than larger models, longer contexts, or better prompts.
Core Claim
Iterative retrieval-augmented generation (RAG) means an LLM alternates between reasoning and fetching new documents. It closes much of the multi-hop gap on Wikipedia questions. Interleaving chain-of-thought with retrieval (IRCoT) improved retrieval recall by 11–21 points. Training on filtered retrieval chains (CoRAG) lifted MuSiQue recall from 29.0% to 47.1%. Tree search over sub-questions (GRAT) and dual text-plus-graph retrieval (GraphSearch) push further. These results motivate our work, but they also constrain it. In Wikipedia questions an intermediate answer is a bare entity ("Christopher Nolan"), and entity identity is all the next hop needs.
In sequestration biology the intermediate answer is a claim with conditions attached. Liao et al. show three conditioned facts. First, CCT8 binds wild-type p53 in the cytoplasm and blocks its nuclear entry. Second, the downstream rise in CDK4/6, Cyclin D1 and EMT markers occurs only in p53-wild-type cells, not in p53-mutant HCT15. Third, CCT8 abundance correlates inversely with nuclear p53 in tissue (ρ = −0.316). Surati et al. show c-Met shifting toward the cytoplasm in tumor while staying membrane- and nucleus-localized in normal lung. They also show glucocorticoid receptor total abundance failing to separate tumor from normal while still carrying strong prognostic signal. Cao et al. show that ZSWIM4 suppresses KIT and BMAL1 only when it enters the nucleus, and that a mutant lacking 45 N-terminal residues fails to enter and loses that activity.
A chain that remembers "p53, CCT8" but forgets "wild-type, cytoplasmic fraction" will happily join this evidence to a paper on mutant p53, for example the cytosolic E285K variant in Dunsche et al.. It may produce a plausible but wrong mechanism. This is the multi-hop version of the "right answer, wrong reason" failure that CGBench documents for single-paper variant curation.
The new element is the qualifier record and the rules for using it. Each hop's answer is stored as a small typed claim: protein, compartment, genotype or cell context, perturbation, direction of effect, and the evidence object that supports it (text span, table cell, or figure panel). The next sub-query is generated from the full record, not from the entity alone. A retrieved passage counts as support only if its qualifiers are compatible with the record. When the compartment or genotype slot cannot be filled from text, the system calls a figure-inspection step. All of this is plain LLM reasoning, not code execution. That choice is deliberate: DISSECT shows biology accuracy collapsing by 73–100 points when models must write and run code.
Research Program
The first deliverable is a benchmark, provisionally SeqHop. Each question requires two to four hops across at least two papers on protein mislocalization: sequestration, failed nuclear entry, aberrant nuclear entry, or surface mislocalization. We will start from mechanisms in the literature of the kind summarized above (CCT8–p53, TRAF7–p53 from Wang et al., ZSWIM4–KIT, MBP-1 from Hippner-Kunicka et al., TBX5 from Steimle et al., SFPQ from Yang et al., and nuclear EGFR and ZIP4). We will extend to a few hundred open-access papers on related proteins.
Biologist annotators write each question together with its gold chain. For every hop they record the gold qualifier record and a modality label: does the needed fact appear in prose, in a table, or only in a figure panel? Each question also comes with qualifier traps: retrievable passages that match the entities but differ in compartment, genotype, or perturbation. One example is a whole-cell lysate blot where the question needs the nuclear fraction. Another is a p53-mutant line where the question concerns wild-type p53. Our target is about 400 questions. That size is enough to detect a 10-point accuracy difference between paired systems with a McNemar test.
The second deliverable is the system. The corpus is ingested following Brkić et al.: text, tables converted to Markdown, and figures stored with caption, model description, and axis text as separate retrievable chunks. We keep their answerability reranker, which scores whether a chunk actually contains the evidence rather than whether it is merely on topic. We extend it so that answerability is judged against the current qualifier record. The chain controller follows the IRCoT pattern but emits a qualifier record at each step. The zoom loop, which requests high-resolution crops of figure panels, is triggered only when a required slot is still empty after reading the text and caption. This should cut the 6-minute median per-figure latency reported by Brkić et al., because most hops should not need it.
The third deliverable is analysis. We will measure how often current systems drop qualifiers, at which hop, and from which modality, and whether the figure-reading errors Brkić et al. call "reading failures" are concentrated in compartment judgments.
Baselines
The naive reference point is a closed-book LLM answering from memory, together with one-shot dense retrieval over text chunks feeding the same LLM. This tells us how much of SeqHop is answerable from pretraining knowledge of well-known mechanisms like p53. We will report closed-book accuracy separately on questions built from post-training-cutoff papers.
The strongest multi-hop text baseline is a CoRAG-style chain. On Wikipedia-style benchmarks the file reports a MuSiQue retrieval recall of 47.1% versus 29.0% for single-step retrieval, and an Exact Match gain of more than 10 points over IterDRAG (CoRAG). We run it with the same base model and the same retrieval budget (maximum hops and chunks per hop) over a text-only index of the corpus. IRCoT is run as a training-free counterpart; its reported gains are 11–21 points of retrieval recall and up to 15 points of QA F1 over one-step retrieval (IRCoT).
The strongest multimodal baseline is the literature agent of Brkić et al., which reports 62.5% on FigQA2, 88.8% on TableQA2 and 88.6% on LitQA3 in LAB-Bench 2. Because it was built for single-paper questions, we also run the most dangerous composite: IRCoT driving retrieval over Brkić's multimodal chunk store, with reranker and zoom loop intact. This composite is the real competitor. It differs from our system only in the qualifier record and the rules for using it, so it doubles as the key ablation.
A further ablation keeps the record but drops the compatibility check, which isolates whether gains come from better queries or from rejecting mismatched evidence. A prompt-only control instructs the composite to "state compartment and genotype at every step" without structured slots.
No reported result exists on Lacuna for any of these systems on cross-paper mislocalization questions. Every number on SeqHop will be new. All systems get the same corpus, base model, and hop and token limits, plus a matched tuning budget on a held-out development split. Beating the composite means a statistically significant gain in final-answer accuracy on trap-containing questions, together with a lower rate of chains that cite qualifier-mismatched evidence. Matching it on accuracy while halving zoom calls would also count, as an efficiency result.
Why not just…?
Why not just run text-only multi-hop RAG over abstracts and results sections? This is the most serious threat. Papers often state localization conclusions in prose; Liao et al. say in words that CCT8 keeps p53 out of the nucleus. The per-hop modality labels let us settle it directly. If text-only CoRAG matches the multimodal systems on hops labelled figure-only, either the labels are wrong or the figures are redundant. In that case we will report that figure reading is unnecessary for this task and recast the contribution as qualifier tracking alone. The qualifier-record ablation on the text-only index tests exactly that.
Why not just give the Brkić agent all candidate papers in one long context? Brkić et al. answer within a single paper. Their latency and reading-failure analysis suggests that more figures in context mean more chances to read the wrong panel, and IRCoT reports that piling up passages saturates context. We test this directly with a long-context condition at matched token cost. If it wins, the cascade structure is unnecessary.
Why not just prompt the model to watch for compartments? It is cheap, and if it works it should be preferred. It is included as the prompt-only control. CGBench's finding that models reach correct labels with wrong supporting evidence, even with 30-shot prompting improving explanation agreement, suggests instructions alone do not enforce evidence consistency. Only the experiment decides.
Why not just extract a knowledge graph and query triples, as in GraphSearch? Subject–predicate–object triples like (CCT8, binds, p53) are exactly where compartment and genotype get dropped. We will run a GraphSearch-style triple channel (GraphSearch) and measure its trap accuracy. If triples extended with qualifier fields perform as well, then our record is effectively a graph schema, and we would say so.
Proposed Experiments
The first decisive check is trap accuracy. On questions where a qualifier-mismatched passage is retrievable, does the composite baseline select it, and does the qualifier record prevent that? The second is chain faithfulness, measured CGBench-style: among correct final answers, the fraction whose every cited hop is gold-compatible. The third is cascade robustness, adapted from the early-divergence concern in CoRAG. We inject a wrong first-hop record, for example a mutant genotype, and measure whether the system flags the conflict or drifts. The fourth is a modality ablation that replaces figures with captions only. The fifth is a transfer sanity check on LAB-Bench 2 FigQA2 and TableQA2, to confirm the gated zoom does not reduce single-paper accuracy below Brkić et al.'s reported 62.5% and 88.8%, along with zoom calls and wall-clock time per question.
The hypothesis is falsified if the prompt-only control or text-only chain matches the full system on trap accuracy and faithfulness. In that case qualifier dropping is real but cheap to fix, and the right paper is the benchmark plus that negative result.
References
- Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions
- Chain-of-Retrieval Augmented Generation
- GRAT: Guiding Retrieval-Augmented Reasoning through Process Rewards Tree Search
- GraphSearch: An Agentic Deep Searching Workflow for Graph Retrieval-Augmented Generation
- CCT8 recovers WTp53-suppressed cell cycle evolution and EMT to promote colorectal cancer progression
- Proteomic characterization of non-small cell lung cancer in a comprehensive translational thoracic oncology database
- Entry of ZSWIM4 to the nucleus is crucial for its inhibition of KIT and BMAL1 in gastrointestinal stromal tumors
- A cytosolic mutp53(E285K) variant confers chemoresistance of malignant melanoma
- CGBench: Benchmarking Language Model Scientific Reasoning for Clinical Genetics Research
- Same Biology, Different Scores: Quantifying the Tool-Use Confound in Agentic Biology Evaluation
- Downregulation of ubiquitin E3 ligase TNF receptor-associated factor 7 leads to stabilization of p53 in breast cancer
- Overexpression of c-MYC Promoter Binding Protein-1 Enhances Proliferation and Glucose Metabolism of Melanoma Cells Lines
- Single-nuclei transcriptomics reveals TBX5-dependent targets in a patient with Holt-Oram syndrome
- SFPQ subcellular mislocalisation entails altered cancer cell function
- A Multimodal Literature Agent as Substrate for Autonomous Biology Research