{"id":"335875f6-8619-4ace-b544-b6c7ab0b875c","arxiv_id":"2607.13298","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Entity-centered focus-guided streaming memory lifts Qwen3-VL-8B to 82.0/69.1 Perception/Backward on OVO-Bench and 74.5 on StreamingBench while cutting writer tokens by ~32%.","lead":"FOLIO is a memory system for live-video AI: it keeps detailed notes on important people and objects while compressing the rest, so questions about earlier moments can be answered without storing every frame. On streaming-video QA benchmarks it reports higher accuracy than prior memory methods and about a third lower memory-writer cost.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Backward SOTA headline is dominated by HLD and FOLIO's 'Unable to answer' prompt; excluding HLD, FOLIO's edge over OASIS shrinks from 11.9 to ~0.3 points, so the 69.1 claim is not yet attributable to the memory system.","rationale":"I read the strongest claim as a state-of-the-art claim for streaming video understanding, supported by the two benchmark numbers. For that claim to hold, the comparisons must isolate the memory design from prompt effects. They do not: HLD is both the largest Backward gain and explicitly attributed to conservative answer calibration. The paper says HLD should be reported separately, then includes it in the 69.1 average. Recomputing from Table 1, HLD explains essentially all of the Backward margin, so the headline overstates the contribution of focused semantic memory. This is an internal consistency issue, not a disagreement with the community. I considered writer VLM fidelity as the weakest link; the paper's own failure taxonomy shows missing evidence in 36% of reviewed OVO errors, but that limits rather than invalidates the architecture. The HLD/prompt confound is more load-bearing because it undermines the specific SOTA number in the abstract. A single rerun with the same answer prompt on the OASIS baseline would settle it. If the confound is confirmed, the paper remains plausible as a memory-system contribution but must be conditional on reporting non-HLD and prompt-controlled comparisons. Hence no verdict change from the reader's CONDITIONAL.","tokens_in":33635,"tokens_out":9060,"duration_ms":90472,"concrete_test":"Re-run Qwen3-VL-8B+OASIS on the OVO-Bench Backward/HLD subset using FOLIO's exact Mode A answer prompt (Appendix K.3) while keeping OASIS memory construction; then recompute Backward averages with and without HLD. If OASIS HLD rises from 47.3 toward FOLIO's 84.6 and the non-HLD Backward margin between FOLIO and OASIS remains below ~2 points, the headline 69.1 SOTA is an answer-prompt artifact. Report both full and HLD-excluded Backward averages for all rows of Table 1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing problem is not writer fidelity but the headline Backward number itself. FOLIO's 69.1 Backward average includes HLD (84.6), a task designed around deliberately irrelevant content whose correct answer is typically 'Unable to answer.' Appendix K.3's Mode A answer prompt explicitly instructs the model to select 'Unable to answer' unless memory/frames contain DIRECT, EXPLICIT evidence, and bans speculative wording. The paper itself says HLD 'benefits from conservative answer calibration' and should be reported separately (Section 4.2). On the same backbone, Qwen3-VL-8B+OASIS gets HLD 47.3, while Qwen3-VL-8B+FOLIO gets 84.6. Excluding HLD, FOLIO's Backward average is (59.0+63.6)/2=61.3 versus OASIS's (62.0+60.1)/2=61.05 — a 0.25-point edge, not the 11.9-point margin implied by the headline. The HLD difference contributes roughly 12.4 points of the 11.9-point Backward margin. Thus the state-of-the-art 69.1 claim is largely an artifact of the answer prompt, not of the focused semantic memory. The central accuracy claim needs re-reporting without HLD and with prompt-controlled baselines.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"FOLIO is a training-free streaming-video memory system that writes, at segment granularity, an entity-centered semantic memory guided by a dynamic focus state, alongside a short-term visual buffer and a visual-evidence cache. At query time it uses lightweight hybrid retrieval over the structured ledger, with VLM-assisted semantic expansion when direct matching fails. The paper reports state-of-the-art accuracy on OVO-Bench (82.0 Perception / 69.1 Backward with Qwen3-VL-8B) and StreamingBench (74.5 overall), with reduced writer cost relative to fixed-uniform writing. The main design claims are supported by component ablations, cost analyses, qualitative cases, and verbatim prompt templates.","tokens_in":34014,"tokens_out":3630,"duration_ms":43715,"significance":"If confirmed, FOLIO would be a useful instantiation of entity-centric, focus-guided memory for online video QA, and the paper's combination of structured memory, explicit focus-state updates, and recoverable keyframe evidence is well matched to backward-tracing and multi-turn questions. The manuscript is unusually transparent in several respects: the writer, expansion, and answer prompts are reproduced verbatim; the entity-merging and relevance-scoring equations are given; and the failure taxonomy separates missing memory from retrieval failures. The central quantitative claim, however, is weakened by the treatment of the HLD subtask, and the marginal contribution of the long-term semantic memory over the recent visual window is modest once HLD is discounted. The architectural ideas remain worth publishing after a corrected and more carefully controlled comparison.","major_comments":[{"comment":"The headline Backward claim is dominated by HLD. Qwen3-VL-8B+FOLIO's Backward average of 69.1 includes HLD=84.6; the paper itself states that HLD 'benefits from conservative answer calibration' and should be reported separately, yet Table 1 includes it in the average and no HLD-excluded row is given. The Mode A answer prompt in Appendix K.3 explicitly instructs the model to select 'Unable to answer' unless memory/frames contain DIRECT, EXPLICIT evidence, and bans speculative wording. Since HLD questions are deliberately irrelevant and the correct answer is typically 'Unable to answer,' the 84.6 HLD score largely measures prompt calibration rather than memory quality. Excluding HLD, FOLIO's Backward average is (59.0+63.6)/2=61.3 versus OASIS's (62.0+60.1)/2=61.05, a 0.25-point margin rather than the claimed +11.9. The state-of-the-art statement and the 'especially clear on Backward querie","section":"§4.2, Table 1, Appendix K.3"},{"comment":"The evidence ablation does not show a large standalone contribution of the long-term semantic memory on the OVO-Bench Backward split. On the 200-query split, short-term visual buffer alone (S) attains 54.5, S+O attains 55.5, and the full FOLIO policy attains 56.7; O alone is 42.6. Thus the structured memory adds roughly 1–2 points over the recent-frame baseline on this split, and the stronger claimed Backward advantage appears only in the HLD-heavy full-table comparison. Because the paper argues that entity-centered memory is the key differentiator over event memories such as OASIS, the authors should report Backward accuracy by subtask with HLD excluded, and should state how much of the remaining gain comes from the structured ledger versus the answer prompt and the recent window.","section":"§4.4, Table 5"},{"comment":"The fixed-hyperparameter claim is not verifiable as written. Equation (1) and Eqs. (5)–(6) introduce coefficients gamma, alpha, beta, retrieval weights w_T,w_H,w_Y,w_V,w_B, entity-matching weights lambda_*, and thresholds tau_m, all described as fixed method hyperparameters, but their values are not given anywhere in the main text or appendices. In addition, the default segment length L=8s is selected using a 200-video StreamingBench diagnostic subset (Fig. 3), and the cost tables (Table 4) use the same 8s setting. Without numerical values and with segment length chosen on an evaluation subset, the claims 'no per-dataset tuning' and the reported cost/accuracy tradeoff are difficult to reproduce and are at risk of diagnostic-subset overfitting. The authors should disclose the hyperparameter values and, if possible, report the sensitivity of the main tables to L.","section":"§3.2, §4.3, Fig. 3"},{"comment":"The system's accuracy depends on the writer VLM producing structured entity/action records that are faithful enough for later queries, but writer fidelity is never measured against ground truth. Table 11 acknowledges that a substantial share of OVO-Bench errors (36%) are 'memory missing' cases where the writer fails to capture fine-grained states, OCR/score changes, or action transitions, and the qualitative cases in Appendix F show similar failures. Since the cost/accuracy advantage of FOLIO is predicated on focused write detail, the paper should either provide a direct fidelity evaluation of writer records (e.g., precision/recall of entity states against annotated keyframes) or explicitly bound the effect of writer misses on the reported accuracy. At minimum, the failure taxonomy should be accompanied by error counts per OVO-Bench subtask, especially EPM and ASI.","section":"§3.4, Appendix K.1, Table 11"}],"minor_comments":[{"comment":"The sentence 'HLD ... should be reported separately' is not operationalized: Table 1 does not include a separate HLD-excluded Backward average. Please add a row or column for Backward minus HLD.","section":"§4.2"},{"comment":"'State-of-the-art' is a strong claim given that Table 1 compares only OASIS among online memory baselines on OVO-Bench, and the comparison excludes the FAR subtasks. Please qualify the claim or add more baselines.","section":"Abstract / Table 1"},{"comment":"Main text refers to 'Appendix 11' for the failure taxonomy; the actual location is Appendix G / Table 11. Please fix the cross-reference.","section":"§4.4 / Appendix G"},{"comment":"Mode A's banned-word list ('most likely', 'suggests', 'implies', 'could be', 'probably') is a strong prior toward abstention. This should be acknowledged as a design choice that can help HLD while potentially hurting other Backward tasks, and its effect should be measured.","section":"Appendix K.3"},{"comment":"Writer latency of 5.88–7.68s per 8s chunk means the writer consumes most of the segment budget. The paper mentions batching as future work, but a sentence connecting this to real-time compliance (or lack thereof) would clarify the streaming claim.","section":"Table 4"}],"recommendation":"major_revision","confidential_remarks":"The paper has real strengths: a coherent entity-centered memory design, transparent prompt templates, and useful cost/failure analyses. My main concern is the HLD/answer-prompt confound, which inflates the headline Backward number and the associated state-of-the-art claim. The authors should be asked to re-run or re-report with HLD excluded and, ideally, with OASIS and the backbone given the same conservative answer prompt. If the memory-specific gain remains small after this correction, the contribution may still be publishable as a systems/memory-design study, but not with the current 'state-of-the-art' framing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real system contribution, and it is also the kind of paper where the authors’ own caveat undercuts their headline. The entity-centered memory unit — persistent object ledger with focus-guided write detail and a recoverable keyframe cache — is a genuine departure from token/KV and event/scene memory families. The cost diagnostics are careful, and the ablations are refreshingly honest: removing the recent window costs 38 points, and they openly say the short-term visual buffer is indispensable. That is a useful boundary result, not a flaw. The multi-turn focus feedback is also a solid, novel mechanism, and the per-component ablation gives it real support. If I worked on streaming video memory, I would want to read this paper and build on the ledger idea. The paper deserves a serious referee. But the central accuracy claim as written is not supported. The 69.1 Backward average is dominated by HLD, and the paper itself says HLD benefits from conservative answer calibration and should be reported separately — then averages it back in. The Mode A answer prompt explicitly forces “Unable to answer” unless direct evidence exists, so HLD is a test of the prompt as much as the memory. Excluding HLD, the FOLIO edge over OASIS on Backward shrinks from 11.9 points to about 0.25 points. That is not a state-of-the-art result. This is the single most important correction: re-report Backward without HLD, run baselines with the same strict-answer prompt, and let the perception numbers (3.9 points over OASIS on 82.0 vs 78.1) carry the accuracy claim instead. Other soft spots are smaller but real: no code or data, most hyperparameters undisclosed (focus coefficients, matching weights, thresholds), no error bars, and the 8s segment length is chosen on a 200-video subset without a stated disjointness guarantee. The writer fidelity issue is honestly acknowledged in the failure taxonomy — fine-grained states, OCR/score changes, and action transitions are missed — but they never measure record fidelity against ground truth, so the accuracy/cost tradeoff rests partly on an unverified assumption. The citation pattern is normal for the area; self-citations are not the issue. Who is this for? Researchers building memory modules for video agents and streaming QA systems. It is a useful design point, and the failure taxonomy gives future work a good map. My recommendation: send it to peer review and push for a revision that separates HLD, controls the answer prompt across baselines, and releases at least the entity ledger code and hyperparameters. Then the paper will be solid.","headline":"A genuinely new entity-centered memory design for streaming video, but the headline Backward score is mostly a prompt artifact and needs to be re-reported before the SOTA claim is fair.","tokens_in":34544,"tokens_out":1916,"would_cite":true,"duration_ms":51020,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Streaming video questions are best answered from a memory organized around persistent entities, written with focus-guided detail, rather than from compressed frames or event summaries alone, and FOLIO shows how to build and maintain that me","keywords":["streaming video understanding","semantic memory","entity tracking","focus state","memory compression","retrieval-augmented video QA","vision-language models","multi-turn video dialogue"],"falsifier":"Build a ground-truth state timeline for a set of streaming videos annotated with scoreboard changes, object locations, and action transitions, then compare FOLIO's written entity records against that timeline. If the records frequently omit score changes or fine-grained action transitions (as the paper's own failure analysis suggests), yet the accuracy figures still hold, then the accuracy gains would not come from the memory records; alternatively, if record fidelity is high but accuracy does not track it, the focus-guided writing claim would need revision. A more direct test: replace FOLIO's","tokens_in":33507,"feed_emoji":"🎬","tokens_out":2515,"duration_ms":32159,"temperature":0.7,"pith_summary":"FOLIO claims that the right unit of memory for streaming video understanding is the persistent entity, not the frame, token, scene, or event. As a video stream arrives, a dynamic focus state decides which entities and actions deserve detailed records and which can be kept compact, so the memory stays small while the information later queries need remains recoverable. The system is training-free: it uses an off-the-shelf vision-language model as a writer that produces structured entity records, merged into a long-term object ledger linked to cached keyframes. On OVO-Bench and StreamingBench, this focused entity memory improves accuracy over prior streaming-memory baselines while cutting writer input, output, and stored-memory size by roughly a third. A sympathetic reader would care because it suggests that memory cost and question-answering accuracy are not necessarily in conflict, if the memory is organized around the objects and actors that questions actually refer to.","feed_headline":"Entity-led memory lifts streaming video QA to 82 percent","feed_subtitle":"FOLIO groups memory around persistent objects and writes focus entities in detail, cutting write cost by a third.","key_machinery":"The central object is the entity-centered long-term semantic memory (the 'object ledger'), which stores each observed entity as a persistent slot with a canonical name, aliases, category, attributes, observation sequences, event sequences, and pointers into a visual-evidence cache. The mechanism that carries the argument is the dynamic focus state, a scalar score per entity updated by a fixed rule p_i(o) = clip(γ p_{i-1}(o) + α·φ⁺_i(o) − β·φ⁻_i(o)), where φ⁺ aggregates visibility, reappearance, state/location change, event participation, and interaction relevance, and φ⁻ aggregates absence and static background behavior. This focus state induces per-segment writing levels that tell the write","core_discovery":"FOLIO shows that an online video memory system can get better streaming question-answering accuracy while writing less, if it writes around entities rather than around time. Its long-term semantic memory is an entity-centered ledger: each observed person, object, or text region gets a stable identity slot with aliases, attributes, a location chain, an action chain, relations, and links to cached keyframes. A focus state, updated per segment from visibility, reappearance, state change, and interaction relevance, assigns each entity a writing level (focus, support, context, or drop) so that important records get fine detail and surrounding context stays compact. At query time, lightweight retr","pith_inferences":["If writer fidelity is the main bottleneck, then a better writer or a schema with domain-specific fields (scores, possession, fine-grained gestures) would amplify FOLIO's gains; the paper's own failure analysis points in this direction.","The entity-ledger idea should transfer beyond video to any streaming perception task where questions reference persistent objects, such as embodied robots or AR assistants, where a compact object history could replace raw sensor logs.","A testable extension is making the writing budget depend on predicted query difficulty or user history rather than a fixed per-segment budget, potentially preserving the cost savings while capturing rare but important state changes.","The ~32% cost reduction was measured on a diagnostic split; scaling to longer streams with many concurrent entities may change the write-detail tradeoff, so the entity-merge step's behavior on streams with hundreds of objects is worth separate study."],"forward_implications":["Long-horizon backward questions, such as where an object was minutes earlier or whether an event ever happened, can be answered from an entity ledger built online, reducing the need to keep raw frames or dense event summaries.","Multi-turn dialogue can steer memory construction: entities mentioned in earlier questions get a focus boost, so later turns about the same targets are grounded in more detailed records.","Focus-guided writing cuts the cost of maintaining streaming memory by around a third, without sacrificing accuracy, meaning memory size need not grow linearly with stream length when only a few entities matter.","Structured entity memory gives a direct way to detect hallucinations: a question about an entity or event that never appears can be answered 'unable to answer' from presence evidence in the ledger.","The method is training-free and works with the same VLM used for writing, retrieval expansion, and answering, so it can be applied to new backbones without fine-tuning."],"fun_headline_variants":["Entity memory for video QA: 82% accuracy, a third less writing","Video QA: entity-focused memory hits 82%, cuts write cost by 1/3","Training-free entity memory streaming video QA to 82% at lower cost","FOLIO: entity-led streaming memory scores 82% QA, writes less","Forget timestamps: entity memory lifts video QA to 82%, slashes writes"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The writer vision-language model, given a few selected keyframes per 8-second segment and a writing-budget instruction, produces entity and action records faithful enough that later queries can be answered from those records.","fun_headline_variants_meta":{"raw":{"variants":["Entity memory for video QA: 82% accuracy, a third less writing","Video QA: entity-focused memory hits 82%, cuts write cost by 1/3","Training-free entity memory streaming video QA to 82% at lower cost","FOLIO: entity-led streaming memory scores 82% QA, writes less","Forget timestamps: entity memory lifts video QA to 82%, slashes writes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001018,"raw_usage":{"total_tokens":4146,"prompt_tokens":768,"completion_tokens":3378,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":3272}},"tokens_in":512,"tokens_out":3378,"duration_ms":26948,"temperature":1.0,"reasoning_tokens":3272,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T05:35:07.427878+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a ground-truth state timeline for a set of streaming videos annotated with scoreboard changes, object locations, and action transitions, then compare FOLIO's written entity records against that timeline. If the records frequently omit score changes or fine-grained action transitions (as the paper's own failure analysis suggests), yet the accuracy figures still hold, then the accuracy gains would not come from the memory records; alternatively, if record fidelity is high but accuracy does not track it, the focus-guided writing claim would need revision. A more direct test: replace FOLIO's","supporting_citations":[],"review_version":1}