{"id":"a019a414-9280-4ce8-aaab-479c1e26f0be","arxiv_id":"2607.02588","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"low","formal_verification":"none","parameter_count":6,"one_line_summary":"Hierarchical online memory with explicit temporal-causal event edges plus a verify-and-correct agentic reasoner yields large gains on long-form video QA in the streaming setting.","lead":"HOMER builds a three-layer online memory (keyframes, entity graph, event graph with causal edges) plus an agent that multi-round retrieves and self-corrects to answer questions on hour-long videos. It beats prior agent baselines by 4–11 points on standard long-video benchmarks and lifts multiple LLM backbones without retraining.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified that overturns the central empirical claim.","rationale":"The paper's central claim is empirical and comparative, not that memory construction is solved. Reader correctly flags the dominant residual error source (§5.4) and the closed-model / multi-knob caveats, yet those are already disclosed and do not reverse the controlled lifts vs. M3-Agent-Gemini3 or the backbone-transfer results. The EventGraph ablation (Table 2) and harness/self-evolution analyses supply the structural evidence the claim needs. A stronger absolute memory writer would raise the ceiling, but is not required for the reported outperformance to be valid. Verdict remains ACCEPT; no adjustment warranted.","tokens_in":19765,"tokens_out":466,"duration_ms":4626,"concrete_test":"Re-run the M3-Bench-robot ablation of Table 2 with an oracle EventGraph (human- or offline-full-video-corrected segments and edges) versus the online incremental patches; if the gap between EntityGraph-only and full HOMER shrinks by more than half of the reported +6.7 overall points, the online construction quality would be load-bearing for the headline gains rather than merely residual.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption (EventGraph construction quality) is real and is already quantified by the paper: §5.4 attributes >75–80% of residual errors to memory (absent/coarse), not the agent loop. That does not undercut the strongest claim. The claim is comparative and structural: HOMER beats the previous best agent (M3-Agent-Gemini3, same memorization pipeline and Gemini-3 backbone, lacking EventGraph + harness) by +5.5 / +10.8 / +4.4 on the three benchmarks, and lifts three backbones including text-only ones (Table 3). Table 2 isolates the EventGraph contribution (+Event Nodes then +edges) under controlled conditions. Because the residual bottleneck is shared with the baseline and the ablations still show net gains from the hierarchical structure and harness, the comparative claim holds even if absolute memory fidelity remains imperfect. No internal inconsistency or untested leap in the reported numbers is evident.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"HOMER is an online long-video understanding framework that maintains a three-layer hierarchical memory (Perceptual Buffer of quality-scored keyframes, EntityGraph of character-centric episodic/semantic memories, and EventGraph of narrative segments linked by explicit temporal and causal edges) and answers queries via an agentic multi-round retrieval loop supervised by a verify-and-correct harness plus a self-evolution skill library. On M3-Bench-robot, M3-Bench-web, and Video-MME-Long it reports +5.5, +10.8, and +4.4 points over the strongest prior agent baseline (M3-Agent-Gemini3, which shares the same memorization pipeline and Gemini-3 backbone but lacks EventGraph and harness). Ablations (Table 2) isolate EventGraph contributions; backbone swaps (Table 3) show gains on three LLMs including text-only models; failure attribution (§5.4) attributes most residual errors to memory construction rather than the agent loop.","tokens_in":20167,"tokens_out":1081,"duration_ms":8852,"significance":"If the reported gains hold under independent reimplementation, HOMER supplies a concrete, model-agnostic structural recipe for bringing multi-hop narrative reasoning into the online (streaming, bounded-memory) regime that prior online methods largely left to the LLM. The combination of typed temporal-causal edges, multilevel retrieval modes, harness engineering, and self-evolution is a useful engineering contribution for embodied and agentic video systems. Strengths that should be credited: transparent failure analysis, controlled memory-layer ablations, backbone-agnostic lifts (including open-source text-only models matching a closed multimodal baseline), and public code. The residual bottleneck the authors themselves quantify (memory construction) is a clear, falsifiable direction for follow-up rather than a hidden flaw.","major_comments":[{"comment":"§4.1 / Table 1: The strongest baseline comparison (M3-Agent-Gemini3) shares the memorization model and Gemini-3 backbone, which cleanly isolates EventGraph + harness. However, EventGraph construction itself is performed by Gemini-3-Flash-Preview (§3.3, §4.1). The paper should report at least one controlled run in which EventGraph construction uses a weaker or open model (or a non-LLM heuristic) so that the structural claim is not confounded by the same strong closed model writing both the graph and the answers. Without that, the model-agnostic claim is only partially supported for the reasoning agent (Table 3), not for the full pipeline.","section":null},{"comment":"§5.4 and Figure 3: >75–80% of residual errors are attributed to Memory (absent) or Memory (coarse). This is a load-bearing limitation for the central claim that hierarchical memory enables reliable multi-hop reasoning. The manuscript should quantify EventGraph edge/segment fidelity (e.g., precision/recall of causal edges against a small human-annotated subset, or agreement under different update cadences) rather than only reporting downstream QA accuracy. Otherwise it remains unclear how much of the +Event Graph lift in Table 2 is genuine causal structure versus better coarse indexing.","section":null}],"minor_comments":[{"comment":"Table 2 vs. Table 1: Memory ablation is on a 600-question subset of M3-Bench-robot while main results use the full split; please state the subset sampling procedure and, if feasible, report the same incremental stack on the full robot split for direct comparability.","section":null},{"comment":"§3.1 Eq. (1)–(2): τ_scene, w, n=3, and δ_min are free parameters; a short sensitivity note (or default values used for all experiments) would improve reproducibility.","section":null},{"comment":"§5.2 / Table 3: Deepseek-v4-Flash and Qwen3-Coder-480B are text-only and never see keyframes; the text should explicitly note that KEYFRAME mode is disabled for those runs so readers do not over-interpret the Perceptual Buffer contribution under those backbones.","section":null},{"comment":"Figure 1 and Algorithm 1 are clear; the harness case study (Table 4) is valuable but dense—consider moving the full multi-round transcript to the appendix and keeping a shorter schematic in the main text.","section":null},{"comment":"Related work: concurrent PyraVid and MM-Mem are cited; a one-sentence contrast on agentic planning vs. automatic graph expansion would help position the contribution more sharply.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The empirical package is solid and the comparative claim against M3-Agent-Gemini3 is the right control. The main risk for a top venue is heavy closed-model dependence for EventGraph construction; requiring one open-model or ablated construction run would substantially strengthen the paper without changing its scope. Fit for a systems/MLLM venue is good."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a clean systems paper that takes M3-Agent’s entity memory and adds what was missing for multi-hop narrative questions—an EventGraph with typed temporal/causal edges, four-mode multilevel retrieval, and a verify-and-correct harness with light self-evolution. The gains are real and not just backbone magic: +5.5 / +10.8 / +4.4 over M3-Agent-Gemini3 on robot, web, and Video-MME-Long, and the same structure lifts Gemini-2.5, Deepseek, and Qwen3-Coder, including text-only backbones that never see keyframes.\n\nWhat is actually new is treating causal and temporal links as first-class retrieval objects rather than leaving the LLM to reconstruct narrative at query time, plus the harness (inform / constrain / verify / correct) and the task ledger that keeps multi-round search from drifting. Table 2 isolates event nodes then edges; Table 3 supports the model-agnostic claim; §5.4 is unusually honest that 75–80% of remaining errors are still memory construction (absent or coarse facts), not the agent loop. That matches the stress-test note: the weakest assumption is EventGraph fidelity, and the paper already quantifies it without hiding behind the agent.\n\nSoft spots, in proportion: heavy closed Gemini dependence for EventGraph construction and reasoning; a pile of free knobs (update cadence, budgets, skill thresholds); self-evolution is batch-level pattern mining on the same distribution, useful but not a deep learning result. Concurrent hierarchical memories (PyraVid, StreamForest, MM-Mem) exist, so novelty is the combination and the agentic control, not the pyramid idea alone. None of that overturns the comparative claim against the shared-backbone baseline.\n\nWho it is for: people building online / streaming long-video agents who care about retrieval structure more than another compression trick. Math is light (systems design, not theory); citations look fair; code is promised. I would bring it to reading group, cite the architecture when discussing online memory, and send it to peer review. Expect referees to push on open re-implementation and denser episodic writers—the authors already point there.","headline":"Solid systems paper: hierarchical memory with causal edges plus a harnessed agent actually moves online long-video QA, with honest ablations and a residual bottleneck the authors already name.","tokens_in":20717,"tokens_out":560,"would_cite":true,"duration_ms":10172,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A three-layer online memory with explicit causal event links, walked by a self-correcting agent, lifts multi-hop answers over hour-long video streams without keeping the full video.","keywords":["long-form video understanding","online memory","hierarchical memory","agentic reasoning","event graph","entity graph","multi-hop reasoning","multimodal large language models"],"falsifier":"Replace the learned Event Graph with human-annotated segments and causal edges on the same videos and re-run the identical agent and harness; if accuracy does not rise substantially, structured event memory is not the main driver. If most remaining errors still fall in “memory absent or coarse” even with oracle graphs, the claim that construction—not the agent loop—is the bottleneck fails.","tokens_in":20686,"feed_emoji":"🎬","tokens_out":956,"duration_ms":19865,"temperature":0.7,"pith_summary":"Models that handle short clips still fail when video arrives as a continuous stream and full history cannot be stored. Prior online systems either keep compressed visual features that lack structure, or higher-level notes ordered only by time, so multi-hop story logic must be rebuilt by the language model at every question. HOMER instead stores raw keyframes, character-centered facts, and event segments joined by explicit temporal and causal edges, then lets an agent locate the scene, drill into details, and verify each step under a harness that corrects failures and accumulates reusable skills. The design produces large accuracy gains on long-video benchmarks that hold when the reasoning backbone is swapped, showing the improvement comes from the memory structure itself rather than any single model.","feed_headline":"Hierarchical memory lifts hour-long video agents by 11 pts","feed_subtitle":"Causal event links and a self-correcting agent enable multi-hop answers under streaming memory limits.","key_machinery":"HOMER’s hierarchical online memory—Perceptual Buffer of keyframes, Entity Graph of character-centric episodic and semantic memories, and Event Graph of segments linked by temporal and causal edges—navigated by an agent that decomposes questions, retrieves at progressive granularity, and is supervised by a verify-and-correct harness with cross-question skill evolution.","core_discovery":"HOMER establishes that online long-video understanding improves when memory separates narrative coverage from perceptual detail and treats temporal-causal relations between events as first-class objects, and when the reasoner can detect insufficient retrieval and recover. With this hierarchical memory and harnessed multi-round agent, the method outperforms the previous best agent baseline by 5.5, 10.8, and 4.4 points on M3-Bench-robot, M3-Bench-web, and Video-MME-Long, and the same structural lift appears across three different language-model backbones.","pith_inferences":["Because memory construction dominates residual errors, better online event segmenters and causal edge writers may yield larger gains than further agent-loop tuning.","The inform–constrain–verify–correct harness is portable to other streaming modalities that require multi-hop evidence under a hard context budget.","Character-centric entity graphs plus causal event graphs form a lightweight narrative schema that could transfer to long-horizon robot or world-model memory.","Skill promotion from trajectories without weight updates offers a practical path for agents that keep improving on a fixed video corpus after deployment."],"forward_implications":["Explicit event-level temporal and causal edges move multi-hop narrative chaining out of the language model’s ad-hoc reconstruction at query time.","The same memory structure lifts both multimodal and text-only backbones, so the gain is structural rather than model-specific.","Failure attribution shows most remaining errors arise in what is written into memory, directing further work at denser online episodic construction.","Batch-level self-evolution produces a compact skill library that improves later questions without retraining any model weights.","Online agents can obtain offline-style multi-hop localization while only ever consulting a bounded, incrementally built memory."],"fun_headline_variants":["Hierarchical memory and agentic reasoning boost long-video agents","Causal event memory lifts hour-long video QA by up to 11 points","Homer: multi-scale memory enables multi-hop answers on streaming video","Online agents gain with hierarchical narrative memory and self-correction","Separating detail from causal events improves long-form video understanding"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The system assumes an LLM can incrementally write accurate event segments and temporal-causal edges from streaming clip notes; if those summaries or edges are wrong or missing, multi-hop retrieval has nothing reliable to walk.","fun_headline_variants_meta":{"raw":{"variants":["Hierarchical memory and agentic reasoning boost long-video agents","Causal event memory lifts hour-long video QA by up to 11 points","Homer: multi-scale memory enables multi-hop answers on streaming video","Online agents gain with hierarchical narrative memory and self-correction","Separating detail from causal events improves long-form video understanding"]},"model":"grok-4.5","effort":"low","cost_usd":0.005194,"raw_usage":{"total_tokens":1479,"prompt_tokens":824,"num_sources_used":0,"completion_tokens":92,"cost_in_usd_ticks":51940000,"prompt_tokens_details":{"text_tokens":824,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":563,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":824,"tokens_out":92,"duration_ms":5218,"temperature":1.0,"reasoning_tokens":563,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T09:30:21.282978+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Replace the learned Event Graph with human-annotated segments and causal edges on the same videos and re-run the identical agent and harness; if accuracy does not rise substantially, structured event memory is not the main driver. If most remaining errors still fall in “memory absent or coarse” even with oracle graphs, the claim that construction—not the agent loop—is the bottleneck fails.","supporting_citations":[],"review_version":1}