When a language model answers a question from a Wikipedia table plus the passages it links to, most of what lands in the prompt is noise: extra rows, off-topic sentences, and garbled text from parsed PDFs. Current systems prune this material by graph distance. They keep everything within three hops of the question's entities, whether or not it bears on the question. This proposal tests a different selection rule. Break every table cell and every passage sentence into its words. Merge all of them into one word co-occurrence graph, where each distinct word is one node and words appearing near each other are connected. Then keep only the evidence whose words fall inside the densest, most tightly interlinked region of that graph that also contains the question's words.
That region is a k-core: the largest subgraph in which every node has at least $k$ neighbors inside it. Earlier work found that k-cores filter out speech-recognition errors better than PageRank does. The core claim here has two parts. First, a query-anchored k-core over a shared word graph is a training-free noise filter that holds up better than distance-based pruning or relevance reranking as the context gets noisier. Second, because a table cell and a passage that share words automatically share nodes, the word graph connects the two sources even where the entity linker (the component that matches cell mentions to Wikipedia pages) fails.
Core Claim
The work builds on three findings in the literature. Combining Graph Degeneracy and Submodularity for Unsupervised Extractive Summarization builds a sliding-window graph-of-words, scores each word by the sum of its neighbors' core numbers (CoreRank), and beats PageRank on noisy meeting transcripts (ROUGE-1 F1 35.88 on AMI, 30.34 on ICSI). The same paper loses slightly to PageRank on clean news (DUC2001). That result both motivates this proposal and constrains it: degeneracy should help most where text is corrupted, and this proposal predicts exactly that.
Hybrid Graphs for Table-and-Text based Question Answering using LLMs (ODYSSEY) supplies the setting. It builds a graph of table cells and hyperlinked passages, prunes by breadth-first search to three hops, and prompts the model hop by hop, reporting up to 10% Exact Match gains and 45.5–53% fewer tokens. ODYSSEY treats passages as monolithic chunks and uses no measure of how cohesive the kept nodes are. A Systematic Exploration of Knowledge Graph Alignment with Large Language Models in Retrieval Augmented Generation shows that centrality predicts which knowledge-graph entities help a model (7.3% average gain). It works only on pure knowledge graphs, and its central nodes are chosen without reference to the question.
The idea here differs from all three in one specific way. Selection is a community search problem: find the maximal cohesive word subgraph that contains query nodes. The search runs over a unified graph whose text side is decomposed into words rather than kept as chunks. The unit that reaches the prompt is still a whole cell or sentence. This matters because answers in QA are often peripheral tokens, such as a date or a number. Whole units are kept if they touch the query's core, so the answer token itself does not have to be central.
Research Program
The graph has three node types: table cells, passage sentences, and normalized word types. Numbers and dates are kept as atomic word nodes, never stemmed. Edges come from three sources. Row and column adjacency connects cells. Containment edges connect each cell or sentence to its words. Sliding-window co-occurrence edges connect words, using window size 2 and 4, following the window-size findings in Graph Neural Networks on Discriminative Graphs of Words.
The candidate pool is whatever a first-stage retriever returns. For Hybrid-QA this is the question's table and its linked passages. For OTT-QA it is the top retrieved fused blocks.
Let $W$ be the word subgraph of the pool and $Q$ the question's content words. Let $k^\*$ be the largest $k$ such that the connected component of the $k$-core of $W$ containing some $q \in Q$ is non-empty, and call that component $C_Q$. Each evidence unit $u$ (a cell or sentence) receives
$$ s(u) = \frac{1}{|u|}\sum_{w \in u} \mathbb{1}[w \in C_Q]\, c(w), $$
where $|u|$ is the number of words in $u$ and $c(w)$ is the CoreRank score of $w$ computed inside $C_Q$.
Units are then chosen greedily under a fixed token budget. The objective is the coverage-plus-diversity submodular objective reused from the CoreRank summarization paper, so that one sub-topic cannot fill the budget. Table rows are serialized with the separator templates that the knowledge-graph alignment study found least harmful. The reader is a zero-shot model. Every reader condition, including the baselines, receives the "return unanswerable" escape instruction from Improving Zero-shot Reader by Reducing Distractions from Irrelevant Documents in Open-Domain Question Answering. This keeps distraction-aware prompting from being confused with the selection rule.
The analysis has two strands. The first is noise-sensitivity curves. Distractor passages and character-level corruption imitating parsing errors are injected at graded rates, and answer accuracy is plotted against noise level for each selector. The second measures linker-free bridging. On the subset of questions where the gold passage is not hyperlinked from any gold cell, or where the linker is deliberately degraded, it checks whether the word layer still recovers the gold sentence.
Baselines
The naive reference points are full context, where the whole table and all linked passages are concatenated, and table-only or passage-only reading. HybridQA: A Dataset of Multi-Hop Question Answering over Tabular and Textual Data reports 8.3% and 19.1% EM for the single-source baselines and 43.8% EM for the supervised HYBRIDER. Full context indicates how much selection matters at all. On UDA FinHybrid, UDA: A Benchmark Suite for Retrieval Augmented Generation in Real-World Document Analysis reports that long-context GPT-4 scores 0.4 without retrieval versus 45.9 with retrieval, which shows how badly unselected context can fail.
The strongest structural competitor is ODYSSEY's breadth-first pruning with hop-wise prompting, because it is the existing zero-shot graph-pruning method for exactly this task. The file gives only relative figures (up to +10% EM over its baseline, −45.5% tokens on Hybrid-QA, −53% on OTT-QA). No absolute EM for ODYSSEY was found on Lacuna, so it will be reimplemented. It will run with the same reader, the same candidate pool, and a token budget matched to the proposed method's.
The strongest reranking competitor is BM25 sentence-and-cell reranking against the question under the same budget. UDA reports BM25 beating dense embeddings on financial documents with exact numbers, so it is a fair and strong choice. A dense cross-encoder reranker will be added as a second variant.
Personalized PageRank seeded at the query words, run over the identical unified word graph, isolates the claim that cohesion (degeneracy) rather than random-walk prestige is doing the work.
Mapping Text to Multiplex Graph: Prompt Compression as Lévy Walk-guided Graph Pruning (RAGP) is the graph-based compression competitor. It reports an average LongBench score of 49.3 at 4x compression versus 48.8 for LongLLMLingua at 3x. It will be run on the flattened candidate pool at the matched compression ratio.
Three ablations isolate the key ingredients. The first removes the word layer, keeping only cell–passage hyperlink edges. The second drops query anchoring, applying CoreRank over the whole pool. The third replaces cores with plain word degree.
All methods get the same tuning budget on a held-out slice of the dev set: 20 configurations each. Beating a baseline means a statistically significant EM gain under a paired bootstrap at equal tokens, or equal EM at clearly fewer tokens.
Why not just…?
Why not just rerank sentences with BM25 or a cross-encoder under the same budget? This is the most serious threat. Query-aware lexical scoring already captures much of what "query-anchored" means. The proposal's claim holds only if two things happen: the core-based selector degrades more slowly as injected noise rises, and it recovers gold evidence on the unlinked-passage subset, where a reranker scoring each unit independently has no way to follow a bridge through a shared cell. If BM25 matches the proposed method on both noise curves and the unlinked subset, the conclusion will be that degeneracy adds nothing beyond lexical overlap in this setting, and it will be reported that way.
Why not just use personalized PageRank from the query? It is simpler and also graph-aware. However, the CoreRank summarization results show PageRank-style prestige failing on corrupted text, where the random walk gets pulled toward generic hub words. The same-graph comparison here tests whether that result carries over. The prediction is a crossover: PageRank equal or better on clean Hybrid-QA passages, cores better on UDA's parsed PDFs and at high injected noise. A crossover would be a useful result in itself, because it would show that the choice of centrality should depend on how noisy the source is.
Why not just feed everything to a long-context model, or compress with RAGP or LLMLingua? UDA's 0.4-versus-45.9 collapse argues against feeding raw full context. ODYSSEY's comparison with summarized prompts argues that summarization erases the exact values multi-hop questions need. Sequence-level compressors such as RAGP are a real competitor, but RAGP is limited to text. Whether it loses on table-heavy questions is measured directly by splitting results into Hybrid-QA's reasoning types, rather than assumed.
Proposed Experiments
The primary benchmarks are the Hybrid-QA dev set in the closed setting, OTT-QA dev in the open setting on top of a fixed first-stage retriever, and UDA's FinHybrid and TatHybrid for naturally noisy parsed documents. On OTT-QA the supervised reference is DoTTeR plus its Cross-Block Reader at 37.8 EM and 43.9 F1 on dev (Denoising Table-Text Retrieval for Open-Domain Question Answering), above the 27.2 EM reported in Open Question Answering over Tables and Text. The goal is not to beat a supervised reader with a zero-shot one. The goal is to narrow the gap without any training, and to check whether that gap narrows further under retriever shift, where Improving Zero-shot Reader by Reducing Distractions from Irrelevant Documents in Open-Domain Question Answering argues supervised readers are fragile.
Metrics are EM and F1, prompt tokens, and gold-evidence recall at a fixed budget: whether the answer-bearing cell or sentence survives selection.
Three checks decide the outcome. First, the noise curves must show the proposed selector's accuracy decaying with a significantly smaller slope than BM25 and breadth-first pruning. Second, on the unlinked-passage subset, the full method must beat the no-word-layer ablation by a clear margin in gold recall. Third, the clean-versus-noisy split must reproduce the direction of the CoreRank-versus-PageRank reversal. If the first two checks both fail, the claim is refuted. If only the bridging check succeeds, the contribution narrows to entity-linker-free fusion, and the note will say so.
References
- Combining Graph Degeneracy and Submodularity for Unsupervised Extractive Summarization
- Hybrid Graphs for Table-and-Text based Question Answering using LLMs
- A Systematic Exploration of Knowledge Graph Alignment with Large Language Models in Retrieval Augmented Generation
- Graph Neural Networks on Discriminative Graphs of Words
- Improving Zero-shot Reader by Reducing Distractions from Irrelevant Documents in Open-Domain Question Answering
- HybridQA: A Dataset of Multi-Hop Question Answering over Tabular and Textual Data
- UDA: A Benchmark Suite for Retrieval Augmented Generation in Real-World Document Analysis
- Mapping Text to Multiplex Graph: Prompt Compression as Lévy Walk-guided Graph Pruning
- Denoising Table-Text Retrieval for Open-Domain Question Answering
- Open Question Answering over Tables and Text