{"id":"52d3733f-2334-4aab-a452-27003899ceb8","arxiv_id":"2606.30093","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"TIGRAG constructs token co-occurrence graphs for scalable graph-augmented RAG and uses iterative entity-driven retrieval to improve multi-hop QA performance over dense and prior graph methods.","lead":"TIGRAG builds a knowledge graph from simple token co-occurrence counts in documents instead of using LLMs to extract entities. This lets retrieval-augmented generation handle multi-hop questions more efficiently while cutting indexing time and prompt size.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Token co-occurrence may fail to identify semantically meaningful bridging entities for multi-hop paths","rationale":"The reader's weakest assumption matches the load-bearing point exactly: the paper's efficiency claim depends on replacing LLM extraction with co-occurrence while preserving multi-hop performance. Full-text details on entity extraction and graph construction would be needed to test this, but the assumption remains the least secure link between method and experimental outperformance. No other internal inconsistency is evident from the abstract.","tokens_in":1678,"tokens_out":341,"duration_ms":20044,"concrete_test":"Sample 30 multi-hop questions from HotpotQA with known gold supporting facts; for each, build the co-occurrence graph on the corpus, run the iterative retrieval, and measure whether the gold bridging facts appear in the top-10 retrieved contexts connected via reported entities. If recall of gold chains falls below 60% while dense retrieval baseline exceeds it, the co-occurrence assumption does not hold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on the token co-occurrence graph (built via sliding-window statistics) plus iterative entity-driven expansion being sufficient to retrieve interconnected evidence for multi-hop QA. The method extracts 'bridging entities' from retrieved contexts to expand queries, but these appear to be surface tokens rather than LLM-derived entities/relations. Co-occurrence captures frequency-based proximity, not necessarily the directed or typed relations required to traverse gold reasoning chains (e.g., in HotpotQA). If this topological model misses or adds spurious links, the reported gains over dense retrieval and LLM-graph baselines would not generalize, as the efficiency advantage would come at the cost of reasoning capability.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes TIGRAG, an efficient graph-augmented RAG framework that constructs a token co-occurrence knowledge graph via sliding-window statistics (avoiding LLM-based entity/relation extraction) and uses iterative entity-driven query expansion with bridging entities plus neural reranking for multi-hop QA. It claims consistent outperformance over dense retrieval and prior graph-based RAG methods on three multi-hop QA benchmarks, together with reductions in indexing time, inference latency, and prompt footprint.","tokens_in":1836,"tokens_out":430,"duration_ms":27043,"significance":"If the experimental claims hold, the work would demonstrate that simple frequency-based token co-occurrence graphs can substitute for more expensive LLM-driven graph construction in multi-hop retrieval, offering a scalable alternative with lower overhead. This would be a meaningful efficiency contribution in the RAG literature, provided the topological model proves sufficient for the required reasoning paths.","major_comments":[{"comment":"Abstract: the central claim that TIGRAG 'consistently outperforms dense retrieval and graph-based RAG methods in both retrieval and downstream QA tasks' is presented with no quantitative metrics, error bars, dataset statistics, baseline implementations, or ablation results. This absence is load-bearing for the experimental contribution and prevents verification of whether post-hoc choices affect the reported gains.","section":"Abstract"},{"comment":"Method description (iterative entity-driven retrieval strategy): the bridging entities used for progressive query expansion are extracted from retrieved contexts via the token co-occurrence graph, yet the manuscript provides no evidence that surface-level co-occurrence links capture the directed or semantically typed relations needed to traverse gold multi-hop chains (e.g., HotpotQA). This assumption is central to the claim that the method achieves both efficiency and reasoning capability without LLM-based extraction.","section":"Method"}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":"The abstract's lack of any numerical results or experimental details is unusually sparse even for an arXiv preprint; this alone would normally trigger a request for the full experimental section before a detailed review can proceed."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the thoughtful comments on the abstract and method. We address each point below and will incorporate revisions to improve clarity and substantiation of our claims.","responses":[{"response":"We agree that the abstract would benefit from quantitative support. In the revised version we will include key metrics (e.g., average retrieval recall@10 and QA F1 improvements over dense and graph baselines on the three benchmarks) together with a brief note on dataset sizes and the main baseline implementations.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the central claim that TIGRAG 'consistently outperforms dense retrieval and graph-based RAG methods in both retrieval and downstream QA tasks' is presented with no quantitative metrics, error bars, dataset statistics, baseline implementations, or ablation results. This absence is load-bearing for the experimental contribution and prevents verification of whether post-hoc choices affect the reported gains."},{"response":"The co-occurrence graph is intentionally undirected and frequency-based rather than semantically typed. The iterative retrieval procedure relies on statistical connectivity to surface bridging entities that empirically connect multi-hop evidence; results on HotpotQA and the other benchmarks indicate that these paths suffice for the required reasoning. We will revise the method section to clarify this design choice, add a short illustrative example of bridging-entity expansion, and note the absence of explicit directionality as a deliberate efficiency trade-off.","revision_made":"partial","referee_comment":"[Method] Method description (iterative entity-driven retrieval strategy): the bridging entities used for progressive query expansion are extracted from retrieved contexts via the token co-occurrence graph, yet the manuscript provides no evidence that surface-level co-occurrence links capture the directed or semantically typed relations needed to traverse gold multi-hop chains (e.g., HotpotQA). This assumption is central to the claim that the method achieves both efficiency and reasoning capability without LLM-based extraction."}],"tokens_in":1362,"tokens_out":412,"duration_ms":16997,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core idea is to build the knowledge graph from raw token co-occurrence counts inside a sliding window instead of paying for LLM calls to extract entities and relations. They then do iterative retrieval that pulls bridging entities from the first round of chunks to expand the query. This targets the indexing-time and latency problems that hit other graph RAG systems.\n\nWhat stands out is the explicit focus on keeping construction cheap and avoiding the error-prone LLM extraction step that most prior graph methods rely on. The iterative entity-driven expansion is a concrete mechanism that tries to recover multi-hop paths without dense embeddings alone.\n\nThe main weakness is that the abstract only asserts consistent wins on three multi-hop QA benchmarks and lower indexing/inference costs, with zero numbers, error bars, dataset sizes, or ablation tables. That makes it impossible to judge whether the co-occurrence graph actually finds the right bridging tokens or just adds noise. The stress-test worry about missing semantic depth looks plausible until the full experiments are checked.\n\nThis is aimed at practitioners who already run RAG pipelines and want lower build costs for connected evidence. It is worth sending to referees if the full paper contains reproducible numbers and fair baseline comparisons; otherwise the claims are too thin to evaluate.","headline":"TIGRAG swaps LLM entity extraction for sliding-window token co-occurrence to build cheaper graphs for multi-hop RAG, but the abstract supplies no numbers so the performance claims stay untested.","tokens_in":2341,"tokens_out":328,"would_cite":false,"duration_ms":14787,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Token co-occurrence graphs let RAG retrieve connected evidence for multi-hop questions without LLM-based entity extraction.","keywords":["retrieval-augmented generation","graph-based RAG","token co-occurrence","multi-hop question answering","knowledge graphs","iterative retrieval","efficient indexing"],"falsifier":"A multi-hop QA test set in which correct answers depend on rare semantic connections that appear infrequently in token co-occurrence counts; if TIGRAG retrieval precision drops below that of LLM-extracted graph baselines on this set, the central efficiency claim would be falsified.","tokens_in":2604,"feed_emoji":"","tokens_out":663,"duration_ms":20691,"temperature":0.7,"pith_summary":"The paper presents TIGRAG, a retrieval-augmented generation system that builds a knowledge graph directly from token co-occurrence counts collected inside sliding windows of text. This graph supports an iterative retrieval process that expands queries by pulling in bridging entities from prior results, combined with neural reranking. The method targets the cost and error issues of prior graph RAG approaches that rely on large language models to pull out entities and relations. Experiments on three multi-hop QA benchmarks show gains in retrieval quality and final answer accuracy alongside lower indexing time, inference latency, and prompt length compared with dense retrieval and other graph methods.","feed_headline":"Token co-occurrence graphs cut RAG indexing time while raising multi-hop QA scores","feed_subtitle":"TIGRAG builds graphs from sliding-window statistics and uses iterative entity expansion to beat dense and LLM-graph baselines on retrieval a","key_machinery":"The token co-occurrence Knowledge Graph, which records direct statistical links between tokens via sliding-window counts to support scalable construction and iterative bridging-entity expansion during retrieval.","core_discovery":"TIGRAG constructs a token-induced graph from sliding-window co-occurrence statistics to model topological relationships between tokens, then uses graph-based semantic expansion together with an iterative entity-driven retrieval loop and neural reranking to surface interconnected evidence; this pipeline outperforms both dense retrievers and LLM-dependent graph RAG systems on multi-hop QA while cutting indexing time, latency, and prompt size.","pith_inferences":["The same co-occurrence graph could be reused across multiple queries without re-extraction, lowering amortized cost for high-volume RAG deployments.","Domains with highly repetitive token patterns may see larger gains than domains with sparse or idiosyncratic phrasing.","Replacing the neural reranker with a purely graph-based scorer could further reduce inference cost while preserving the core topology signal."],"forward_implications":["Indexing time drops because graph construction avoids any LLM calls for entity or relation extraction.","Inference latency and prompt footprint shrink due to smaller retrieved context sets from targeted graph expansion.","Retrieval recall for multi-hop questions rises by iteratively adding bridging entities found in the co-occurrence graph.","Downstream QA accuracy improves on benchmarks that require chaining evidence across documents."],"fun_headline_variants":["Token co-occurrence graphs cut RAG indexing time","Token graphs raise multi-hop QA scores","Co-occurrence graphs lower RAG latency in QA tasks","Token graphs aid multi-hop RAG retrieval efficiency"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Token co-occurrence statistics alone are sufficient to capture the topological relationships needed for effective multi-hop reasoning without the semantic depth provided by LLM-based entity or relation extraction.","fun_headline_variants_meta":{"raw":{"variants":["Token co-occurrence graphs cut RAG indexing time","Token graphs raise multi-hop QA scores","Co-occurrence graphs lower RAG latency in QA tasks","Token graphs aid multi-hop RAG retrieval efficiency"]},"model":"grok-4.3","cost_usd":0.006329,"raw_usage":{"total_tokens":2965,"prompt_tokens":652,"num_sources_used":0,"completion_tokens":56,"cost_in_usd_ticks":63287000,"prompt_tokens_details":{"text_tokens":652,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2257,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":652,"tokens_out":56,"duration_ms":16820,"temperature":1.0,"reasoning_tokens":2257,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-30T06:10:33.966553+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A multi-hop QA test set in which correct answers depend on rare semantic connections that appear infrequently in token co-occurrence counts; if TIGRAG retrieval precision drops below that of LLM-extracted graph baselines on this set, the central efficiency claim would be falsified.","supporting_citations":[],"review_version":1}