{"id":"a90002f7-e0ac-40d4-80c8-feaec2399e15","arxiv_id":"2608.07458","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"CoinRAG reuses offline-computed, nugget-granular KV caches to make retrieval-augmented generation faster and more accurate under a 100 ms prefill budget.","lead":"CoinRAG precomputes fine-grained 'nugget' caches for document chunks and stitches the query-relevant ones together at inference time, skipping full-context encoding. The result is a better accuracy versus latency trade-off for long-context question answering under a 100 ms response budget.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 5.3% average F1 gain is selected from per-dataset hyperparameter sweeps on 200-question eval sets with no significance testing; the 0.5% gap on 2WikiMQA makes the headline number statistically fragile.","rationale":"The reader's weakest_assumption targets the technical question of whether sliced KV caches with rotated positions behave like a fresh composed encoding. That is a genuine limitation, but it is shared in spirit with TurboRAG-style chunk-level reuse, and the paper partially mitigates it with nugget-aware fine-tuning; even an imperfect approximation can still yield the reported empirical behavior. The more load-bearing issue is the statistical basis of the headline numbers. The 5.3% average is computed after per-dataset selection over a large hyperparameter grid on only 200 questions per dataset, with no significance testing, and the 2WikiMQA improvement is essentially a tie. This makes the central quantitative claim unverified. The reader already assigned CONDITIONAL partly for these reasons, so my concern does not move the verdict; it reinforces the condition that the authors must report significance testing, fixed-configuration results, and ideally larger or held-out evaluation sets before the Pareto-frontier claim can be accepted as robust.","tokens_in":21347,"tokens_out":5588,"duration_ms":55151,"concrete_test":"Compute a paired bootstrap (10,000 resamples) of per-query F1 differences between CoinRAG and TurboRAG at the reported best configurations under the 100 ms budget, using the 200-query eval sets; report 95% CIs for each dataset and for the 3-dataset average. In addition, report the comparison with one fixed (kc,k) configuration (e.g., the median of the per-dataset best configs) rather than per-dataset selection. If the average CI includes 0, or the fixed-config gain falls below 5.3%, the headline claim is not statistically established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Section 5 is a numerical comparison against TurboRAG: +4.7%, +0.5%, and +14.6%, averaging +5.3% under a P99 100 ms budget. Table 3 reports the best configuration for each method after sweeping kc in {1,...,50} and k in {1,...,200} (Appendix E) and selecting the Pareto point with the highest F1 at that budget. Because selection is done per dataset on the same 200-question eval sets used to report the headline, the reported gains are maxima over many configurations rather than estimates for a fixed system. No confidence intervals, paired tests, or bootstrap intervals are provided. On 200 questions, a change of a few predictions moves F1 by multiple points, and the 2WikiMQA gap is 0.5% (42.2 vs 42.4 in Table 3), roughly one question. Thus the average 5.3% gain, driven mainly by MuSiQue (+14.6%), could shrink or reverse with a different sample or a pre-registered configuration. This directly undermines the quantitative Pareto-frontier claim, independent of the plausibility of the KV-slicing approximation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents CoinRAG, a RAG system that avoids online encoding of long retrieved contexts by precomputing per-chunk KV caches offline, extracting fine-grained 'nuggets' as token spans, retrieving the most query-relevant nuggets in two stages, and composing their sliced KV caches into a single prompt prefix using RoPE position rotation, with optional nugget-aware fine-tuning. The central claim is that under a P99 TTFT budget of 100 ms, CoinRAG achieves higher answer F1 than chunk-level KV-cache-reuse baselines on three LongBench multi-hop QA datasets, improving over the strongest baseline TurboRAG by 4.7%, 0.5%, and 14.6% (average 5.3%), while also shortening active context length and improving the accuracy-length Pareto frontier.","tokens_in":21489,"tokens_out":10386,"duration_ms":98762,"significance":"If the empirical claims hold, this is a practically useful contribution to efficient long-context RAG: it directly targets an interactive latency SLA, reports detailed engineering trade-offs (disk storage, I/O, offline extraction cost, fine-tuning cost), includes ablations that isolate each component, and retrains TurboRAG and KVLink under matched data and hyperparameters. The paper is also careful in its limitations section about offline costs, checkpoint coupling, retrieval recall, and the absence of cross-chunk attention. I do not see a circularity problem in the F1 comparisons; however, the headline statistical evidence and the validation of the KV-composition approximation are not yet commensurate with the strength of the claims.","major_comments":[{"comment":"The headline average gain is a comparison of per-dataset maxima, not of fixed systems. Appendix E sweeps kc over 9 values and k over 12 values for CoinRAG (108 configurations), while chunk baselines sweep 9 kc values (CacheBlend additionally sweeps 11 recomputation ratios), and Table 3 reports, for each method, the configuration with the highest F1 at the 100 ms budget on the same 200-question evaluation sets used to report the results. The 2WikiMQA difference is 42.4 vs. 42.2 F1, roughly one question in 200, and no confidence intervals, paired bootstrap intervals, or significance tests are reported. Because Section 5 quantifies gains to one decimal place, please add per-question paired bootstrap CIs and either pre-register the configuration or select it on a held-out validation split; otherwise the central Pareto-frontier claim is not statistically established.","section":"§4.2, Table 3, Appendix E"},{"comment":"The central KV-composition mechanism is not validated against the correct oracle. The statement in §2.4 that sliced KV states are 'identical to what they would be under a fresh encoding of the chunk' is about the original chunk-level cache, not about the composed sequence of nuggets from different chunks; cross-slice attention is absent and RoPE rotation only relocates positions. The Isolated Nugget Encoding ablation in §4.3 is a reasonable component study, but it does not measure the approximation error introduced by composing slices rather than encoding the same nugget sequence fresh in one pass. Please add a direct control that encodes the exact composed nugget sequence in a forward pass (reporting F1 with latency unconstrained, and if feasible under a matched budget) and report the difference; this determines whether the mechanism itself is sound or whether the observed gains come from other components such as fine-tuning and prompt design.","section":"§2.4 and §4.3"}],"minor_comments":[{"comment":"The sentence 'it consistently outperforms Standard RAG and Standard CAG with a 5.3% (41.7 vs. 39.6) answer quality improvement' is inaccurate: Table 3 shows 41.7 vs. 39.6 is CoinRAG vs. TurboRAG, while Standard RAG averages 29.9; please correct the sentence or name the right baseline.","section":"§1"},{"comment":"The CoinRAG rows are garbled in the rendered text (e.g., '1, 3051.472 87 618'); please fix the table formatting so each column is legible.","section":"Table 3"},{"comment":"Please state explicitly whether all baselines use the identical answer-generation prompt; if CoinRAG uses a different prompt from the baselines, prompt differences would confound the F1 comparison.","section":"§4.1 and Appendix E"},{"comment":"The paper mixes 'relative improvement' and 'F1 improvement' (5.3% vs. +5.3 points); please use one convention consistently throughout.","section":"§5 and Abstract"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid systems paper and the requested revisions are within scope: add paired bootstrap/CI analysis and the composed-context encoding control. I would not require a new algorithm; the current evidence is insufficient mainly because of per-dataset configuration selection and small evaluation sets. The 0.5% 2WikiMQA gap should not be presented as a meaningful gain without uncertainty quantification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere is my honest read of CoinRAG. The core idea is genuinely new: extract text-span nuggets offline, slice precomputed chunk KV caches at those spans, and stitch the slices together with RoPE position rotation. That combination moves past chunk-level reuse in TurboRAG, CacheBlend, and KVLink, and past online context-free nuggets in GINGER and Crucible. The paper is also unusually transparent: hardware spec, I/O layout, dollar costs, step-by-step inference trace, and honest failure cases. The ablations are informative and each design component earns its place.\n\nThe main soft spot is the headline comparison. The 5.3% average gain over TurboRAG is selected from per-dataset sweeps of kc and k on the same 200-question eval sets, with no confidence intervals or significance tests. The 2WikiMQA gap is 0.5%, which is about one question. So the average reflects controlled experiments, but it is a maximum over configurations, not an estimate for a fixed system. A pre-registered configuration or bootstrap intervals would strengthen the Pareto-frontier claim considerably.\n\nThe second concern is the central approximation. Slicing a chunk's KV cache at a span and rotating positions yields representations computed under the original chunk context. When you concatenate slices from different chunks, the attention patterns inside each slice were never computed under the new composed context. The paper's ablation compares slicing against re-encoding the nugget in isolation, not against a fresh encoding of the composed sequence. So the approximation is plausible but not directly tested. That said, the isolated-encoding ablation shows 4-6 point F1 drops, and the overall gains over chunk baselines suggest the approximation works in practice.\n\nEvaluation is on three LongBench multi-hop QA datasets, with fine-tuning on their training splits. That is narrow but appropriate for the claim. No code or data are released, which limits reproducibility.\n\nBottom line: this is a solid systems contribution that deserves serious peer review. The statistical robustness needs work, and the KV-slice assumption deserves a direct test. I would cite this in related work and recommend conditional acceptance with a request to tighten the statistical claims and release artifacts.\n\nBest,\n[Your name]","headline":"Genuinely new nugget-level KV cache reuse with solid ablations, but the headline 5.3% F1 gain is statistically fragile and the KV-slice approximation is plausible yet under-tested.","tokens_in":22132,"tokens_out":2173,"would_cite":true,"duration_ms":21375,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CoinRAG claims that reusing fine-grained 'nugget' KV caches, rather than full chunks, improves RAG answer quality under a 100 ms latency budget, outperforming the strongest chunk-level baseline by an average 5.3% F1.","keywords":["retrieval-augmented generation","KV cache reuse","long-context question answering","information nuggets","latency-accuracy Pareto frontier","multi-hop QA","prefill latency optimization"],"falsifier":"Run the same selected nuggets through two configurations: (1) CoinRAG's cache-sliced composition, and (2) a fresh forward pass that encodes the concatenated nugget text with the same order and positions, with all other components (retrieval, model, prompt) held fixed. If the fresh encoding attains equal or higher F1 on HotpotQA, 2WikiMQA, and MuSiQue at comparable context length, the claim that cache slicing preserves context worth is falsified; if the fresh encoding is worse, slicing adds value beyond faithful replication.","tokens_in":21047,"feed_emoji":"🪙","tokens_out":7822,"duration_ms":61278,"temperature":0.7,"pith_summary":"CoinRAG claims that the accuracy-latency Pareto frontier of long-context retrieval-augmented generation (RAG) can be improved by operating on fine-grained 'information nuggets' rather than full text chunks. It precomputes key-value (KV) caches for each document chunk offline, extracts nuggets as token spans, and at inference time retrieves query-relevant nuggets and slices the cached KV tensors for those spans, stitching them into a compact prefix with rotary position alignment. Under a strict 100 ms P99 time-to-first-token budget, this yields an average 5.3% relative F1 improvement over the strongest chunk-level cache-reuse baseline across three multi-hop QA datasets, with a roughly 1.8x shorter active context. The design shifts the bottleneck from online prefill computation to one-time offline corpus encoding, and the paper argues that the gain comes from removing noise and redundancy while preserving document-level conditioning through cache slicing.","feed_headline":"Nugget cache reuse lifts RAG F1 by 5.3% at 100 ms","feed_subtitle":"CoinRAG assembles query-relevant text-span caches to cut prefill latency and noise, shifting the accuracy-latency frontier.","key_machinery":"The central mechanism is the contextualized nugget KV cache: each document chunk is encoded once in a full forward pass, and each extracted nugget is stored as a token span [si, ei] within that chunk's cached key-value tensors. At inference, the selected nugget caches are sliced as C[si:ei], rotated by an offset Δi that reassigns contiguous positions via the RoPE rotation operator, and concatenated into a single prefix cache. This lets the model attend to a compact sequence of nuggets as if it were a fresh continuous prefix, without re-encoding any raw text. Two-stage retrieval (first top-kc chunks by dense embedding, then top-k nuggets within those chunks) supplies the query-specific spans, and nugget-aware fine-tuning trains the model on the exact same stitched-cache construction to close the training-inference gap.","core_discovery":"On its own terms, the paper's discovery is that a RAG system can beat chunk-level KV cache reuse on answer quality while staying within a 100 ms P99 TTFT budget by representing retrieved evidence as query-selected nugget spans of a precomputed chunk cache. Concretely, on LongBench multi-hop QA (HotpotQA, 2WikiMQA, MuSiQue), the best CoinRAG configuration reaches F1 scores of 51.4, 44.5, and 32.3 under the budget, against 49.1, 42.2, and 27.4 for the strongest chunk-level competitor, an average improvement of 5.3%. The paper also shows that even with no latency limit, the average F1 gain over all baselines is 5.2%, while the active context length is up to 10.1x shorter than standard RAG. These empirical results are attributed to three mechanisms working together: two-stage chunk-then-nugget retrieval that narrows the candidate pool, context-preserving KV slicing that avoids re-encoding isolated span text, and nugget-aware fine-tuning that aligns the training distribution with the stitched-cache inference structure.","pith_inferences":["The paper's core assumption—that sliced chunk-cache representations behave like fresh encodings of the composed nugget sequence—could be tested directly by comparing CoinRAG's assembled prefix against a fresh forward pass over the same nugget text. If the fresh encoding matches or exceeds F1, the gain would come from context compression rather than from cache-reuse fidelity; if it is lower, slicin","The failure examples suggest an exploitable bias: when supporting and distractor nuggets share surface terms, the model tends to follow frequency. A testable extension is re-ranking selected nuggets with a cross-encoder or applying a frequency-debiasing heuristic at composition time, which could recover some of the missed answers.","Because the offline cache is tied to a specific model checkpoint and positional-encoding topology, the method's practical reach depends on how often the backbone changes. An incremental-re-encoding scheme for changed chunks, noted by the paper, would make the method viable for evolving corpora.","If the slicing approximation holds for other RoPE-based models, the same nugget-cache composition could be applied to long-document summarization or open-domain QA with different retrieval backbones, potentially extending the latency-accuracy frontier beyond the three LongBench tasks."],"forward_implications":["Under a 100 ms P99 TTFT service-level agreement, CoinRAG's best configuration achieves higher F1 than the strongest chunk-level baseline on all three multi-hop QA datasets, with an average relative gain of 5.3%.","The active prefix context is about 1.84x shorter than the best chunk-level baseline under the same budget, which directly reduces the KV cache memory footprint per request and raises the number of concurrent requests a server can hold.","When the latency limit is removed, CoinRAG still improves average F1 by 5.2% over the strongest chunk-level baseline, while using 6.8x shorter context on average, indicating that removing noisy context outweighs the loss from missing cross-chunk interactions.","Position alignment during cache composition contributes most when the latency budget is tight: under a 75 ms P99 budget it improves F1 by 3.0–8.5% across datasets, and becomes comparable to no alignment beyond 100 ms.","Nugget-aware fine-tuning is a load-bearing component: ablating it reduces peak F1 by 6.3 to 11.3 points on the three datasets."],"supporting_citations":[{"why":"TurboRAG, the chunk-level KV cache reuse method that is CoinRAG's strongest competitor; its precomputed chunk-cache design is the baseline that CoinRAG extends to nugget granularity.","marker":"(Lu et al., 2025)"},{"why":"CacheBlend, the selective-recompute baseline that restores cross-chunk attention; provides the comparison for partial recomputation.","marker":"(Yao et al., 2025)"},{"why":"KVLink, the link-token baseline for cross-chunk attention; provides the comparison for trainable cross-chunk linking.","marker":"(Yang et al., 2025)"},{"why":"LongBench, the benchmark that supplies the three multi-hop QA datasets (HotpotQA, 2WikiMQA, MuSiQue) and defines the evaluation setup.","marker":"(Bai et al., 2024)"},{"why":"RoPE position rotation, the mathematical operator that CoinRAG adapts to realign cached nugget positions during cache composition.","marker":"(Su et al., 2024)"},{"why":"BGE-M3, the dense retriever used in both stages of the two-stage nugget retrieval.","marker":"(Chen et al., 2024a)"},{"why":"AutoNuggetizer, the LLM-based nugget extraction approach that motivates the offline nugget representation.","marker":"(Pradeep et al., 2025)"},{"why":"CAG, the cache-augmented generation paradigm CoinRAG builds on, establishing the idea of preloading document KV caches.","marker":"Chan et al. (2025)"}],"fun_headline_variants":["CoinRAG: nugget caches cut RAG latency, lift F1 5.3%","CoinRAG reuses small caches for 5.3% better RAG F1","Fine-grained KV reuse in RAG: 5.3% F1 gain, 10x less context","CoinRAG assembles nugget caches to beat chunk-level RAG","Nugget-level KV cache reuse boosts RAG accuracy and speed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach assumes that slicing a precomputed chunk's key-value cache at an offset span and rotating its positions produces a representation equivalent to freshly encoding the composed nugget sequence in one continuous context.","fun_headline_variants_meta":{"raw":{"variants":["CoinRAG: nugget caches cut RAG latency, lift F1 5.3%","CoinRAG reuses small caches for 5.3% better RAG F1","Fine-grained KV reuse in RAG: 5.3% F1 gain, 10x less context","CoinRAG assembles nugget caches to beat chunk-level RAG","Nugget-level KV cache reuse boosts RAG accuracy and speed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001153,"raw_usage":{"total_tokens":4814,"prompt_tokens":1018,"completion_tokens":3796,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":3680}},"tokens_in":634,"tokens_out":3796,"duration_ms":23839,"temperature":1.0,"reasoning_tokens":3680,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T04:18:34.375101+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same selected nuggets through two configurations: (1) CoinRAG's cache-sliced composition, and (2) a fresh forward pass that encodes the concatenated nugget text with the same order and positions, with all other components (retrieval, model, prompt) held fixed. If the fresh encoding attains equal or higher F1 on HotpotQA, 2WikiMQA, and MuSiQue at comparable context length, the claim that cache slicing preserves context worth is falsified; if the fresh encoding is worse, slicing adds value beyond faithful replication.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"AutoNuggetizer, the LLM-based nugget extraction approach that motivates the offline nugget representation."}],"review_version":1}