{"id":"e820a55c-ad7d-4097-9916-ede3996a90e5","arxiv_id":"2607.10387","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Interleaving periodic time markers with continuous audio tokens, plus duration-mixture synthetic training, yields stable temporal grounding for an audio LLM on inputs up to 120 minutes.","lead":"An open audio language model answers questions about recordings up to two hours long and returns explicit timestamps for the evidence. This design choice makes long meetings, podcasts, and lectures searchable by time rather than by vague summary alone.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Synthetic supervision may teach transcript-pattern matching rather than audio-conditioned temporal localization, undercutting the claim that inter-timings enable genuine long-form grounding.","rationale":"The reader correctly isolates the synthetic-data pipeline as the weakest assumption. The inter-timing and duration-mixture ablations are cleanly reported and internally consistent, yet they inherit the same supervision source; therefore any claim that the anchors enable genuine audio-conditioned localization remains conditional on the unverified transfer from transcript-derived labels. No stronger internal contradiction appears in the manuscript, and the open release of weights/data makes the concrete hold-out test feasible. The verdict therefore stays CONDITIONAL; the stress test simply sharpens the same load-bearing concern rather than introducing a new one.","tokens_in":9948,"tokens_out":511,"duration_ms":9641,"concrete_test":"Hold out a subset of long recordings never seen by WhisperX or the text LLM; generate human-annotated temporal-grounding questions whose answers cannot be recovered from any transcript of that audio; re-evaluate the released checkpoint (and the w/o-inter-timings ablation) on this set. If the mIoU gap between inter=60 s and no-inter collapses or both scores fall near chance, the synthetic-supervision assumption fails and the central claim weakens.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim (Table 1: inter-timings essential; 53.8\to14.2 mIoU collapse without them; Figure 3 duration-mixture necessity) rests on models trained almost entirely on synthetic (Q,A) pairs generated by a text-only LLM from WhisperX-aligned transcripts (Section 3.5, Figure 2). Slice-wise generation + global verifier only enforce consistency against the same transcript; they do not guarantee that the audio encoder must be used for localization. Because the continuous audio tokens are produced from the identical recordings that yielded the transcripts, the model can learn to map acoustic patterns that co-occur with transcript phrases onto the interleaved time markers without performing true temporal reasoning over the audio stream. The short-clip external sets (AudioGrounding, DAQA) are too brief to stress long-form behavior, and AMI questions are phrase-level and themselves auto-generated from transcripts. Consequently the ablation may demonstrate that anchors help the model exploit transcript-derived supervision rather than that they are required for audio-native long-form grounding.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper introduces GigaChat Audio, a 10B-A1.8B MoE audio-conditioned LLM that accepts up to 120 minutes of input and produces answers, fragment descriptions, and summaries with explicit timestamps. The core design interleaves continuous audio embeddings (160 ms frames from an encoder–subsampler–projector stack) with periodic inter-timing markers (plain-text hh:mm:ss or special tokens). Temporal supervision is obtained at scale via a cascaded synthetic pipeline: WhisperX-aligned transcripts, ~10-minute slice-wise (Q,A) generation with a text-only LLM, global verification against the full transcript, and multi-sample aggregation for evaluation. The authors report competitive short-clip results and substantially stronger long-form temporal grounding than open and proprietary baselines, with ablations on marker frequency, format, special tokens, and duration-mixture training (Tables 1–4, Figure 3). Model weights and a 10k+ hour temporal dataset are released.","tokens_in":10274,"tokens_out":1479,"duration_ms":26107,"significance":"Temporal grounding for long recordings is a genuine bottleneck for current audio LLMs; the paper isolates a simple, practical mechanism (periodic inter-timings) and shows, via controlled ablations, that removing anchors collapses 20–40 min mIoU from 53.8 to 14.2 and that duration-mixture training is required for length generalization up to 120 min. The open release of weights and a large temporal dataset is a concrete contribution that can accelerate follow-on work. The engineering choices (marker frequency vs. token cost, plain-text vs. special tokens, silence and language filtering) are carefully quantified and useful to practitioners. If the claimed audio-native localization holds under stronger controls, the work would be a solid reference system for time-aware long-form audio understanding.","major_comments":[{"comment":"Section 3.5 / Figure 2 and Table 1: The central claim that inter-timings enable genuine long-form audio grounding rests almost entirely on synthetic (Q,A) pairs generated by a text-only LLM from WhisperX transcripts of the same recordings used at training time. Slice-wise generation plus a global verifier only enforce transcript-internal consistency; they do not force the model to use the continuous audio stream rather than acoustic co-occurrence patterns that align with transcript phrases and the interleaved markers. AMI questions are themselves auto-generated from transcripts, and the long-form TGr / description / summary sets are synthetic. Please add at least one of: (i) a human-annotated long-form temporal-QA subset, (ii) an audio-ablation or transcript-only control that quantifies residual performance when audio is corrupted or replaced, or (iii) a clear analysis showing that local","section":null},{"comment":"Section 3.6 and Table 1 (Descriptions, Summ columns): Fragment descriptions and timed summaries are scored solely by an LLM-as-a-judge protocol whose correlation with human judgments is not reported for this domain. The paper itself notes that reference-based metrics correlate imperfectly with humans, yet provides no human calibration, inter-annotator agreement, or judge-prompt sensitivity study. Because these scores are used to claim superiority on open-ended timed generation, a modest human validation study (or release of judge prompts + agreement numbers) is needed before the open-ended results can be treated as load-bearing.","section":null},{"comment":"Tables 1–4 and Figure 3: All headline mIoU / MAE / judge scores are point estimates with no error bars, multi-seed variance, or statistical tests. Given that SFT mixtures, synthetic generation temperature, and verifier thresholds are free parameters, single-run differences (e.g., 53.8 vs. 14.2 mIoU, or 60 s vs. 7 s anchors) cannot be assessed for robustness. Please report at least multi-seed standard deviations or bootstrap intervals on the key ablations that support the necessity of inter-timings and duration mixtures.","section":null},{"comment":"Section 4.1 / Table 1: Long-form comparison against strong systems is incomplete. TimeAudio is reported only on short clips (authors note degradation beyond ~2 min); Qwen3-Omni and Gemini 3 Flash are included, but there is no systematic comparison against long-context ASR+LLM pipelines that already emit word-level timestamps (e.g., WhisperX + LLM over the transcript). Such a pipeline is a natural and strong baseline for timestamped QA and would clarify how much of the gain is attributable to end-to-end audio conditioning versus access to aligned time markers.","section":null}],"minor_comments":[{"comment":"Figure 1 caption and Section 3.1: Clarify whether inter-timings are inserted only in the input stream, only in the target, or both, and whether the final end-of-audio marker is always present at inference.","section":null},{"comment":"Table 1 header: “Tm(↑) AES(↑) Rd(↓)” for summarization is dense; expand once in the caption or define AES explicitly as (Acc + 1 − Err + Style)/3 next to the table.","section":null},{"comment":"Section 3.4: The silence-ratio and language-ID thresholds (p(English)>0.7) are free parameters that affect the 14k-hour pool; a short sensitivity note or appendix table would help reproducibility.","section":null},{"comment":"Section 4.3 / Table 2: “extra tokens” require much higher TG data ratios to match plain-text; the discussion could briefly note whether this is an initialization or capacity issue.","section":null},{"comment":"Typos / polish: “V oxtral”, “VibeV oice”, “front-loading bias” hyphenation is inconsistent; “inter-timings” is introduced without a formal definition before Figure 1. A short glossary or first-use definition would help.","section":null},{"comment":"References: Several 2025–2026 arXiv preprints are cited; ensure final versions or DOIs are updated at camera-ready if available.","section":null}],"recommendation":"major_revision","confidential_remarks":"The synthetic-supervision circularity raised by the skeptic is the main load-bearing risk; if the authors can add a human long-form subset or a clean audio-vs-transcript control, the paper becomes a clear accept for an applied audio/ML venue. Scope fit for a serious eess.AS / speech journal is good given the open release and systematic ablations. No concerns about citation manipulation or undisclosed dual submission from the manuscript itself."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing worth knowing is that this is a usable engineering result, not a theory paper. They take a 10B-A1.8B MoE, attach a continuous-audio front-end, and interleave plain-text (or special) time markers every 60 s (or denser). That plus training on a mixture of short-to-long clips gets them stable temporal grounding out to 120 min, with mIoU holding around 53–65 on 20–40 min buckets while the no-anchor ablation falls to 14. They also ship the weights and a 10 k+ hour temporal dataset.\n\nWhat is actually new is the combination of design choices plus the evidence: systematic ablations on marker frequency, format, special tokens vs text, and duration mixtures (Tables 2–4, Figure 3), plus a cascaded synthetic pipeline that slices transcripts to kill front-loading and runs a global verifier. The external short-clip numbers (AudioGrounding, DCASE time-aware, AMI phrase localization) and the head-to-head with Qwen3-Omni and Gemini 3 Flash make the long-form claim concrete. Open release is real credit.\n\nSoft spots are real but proportionate. Almost all long-form supervision is synthetic (Q,A) pairs from a text-only LLM over WhisperX transcripts. The verifier only enforces transcript consistency, so the model could be learning to align acoustic patterns that co-occur with transcript phrases to the interleaved markers rather than doing pure audio-native localization. Short external sets and auto-generated AMI questions do not fully close that loop. Open-ended description/summary scores rest on an unvalidated LLM judge, and the main tables lack error bars or multi-seed runs. Those are addressable gaps, not contradictions of the ablation results.\n\nThis is for people building meeting/podcast tools or long-context audio LLMs. The math is ordinary SFT, the data pipeline is transparent, citations are appropriate. I would send it to peer review; the open artifacts and the clear “anchors + mixture matter” findings are worth referee time even if the synthetic-data concern forces extra experiments.","headline":"Practical open systems paper: sparse inter-timings plus duration-mixture training make hour-scale timestamped audio QA work, with clean ablations and a real release; synthetic-transcript supervision is the main caveat, not a collapse.","tokens_in":10887,"tokens_out":554,"would_cite":true,"duration_ms":20541,"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":"Interleaving periodic time markers with continuous audio tokens lets an audio LLM ground answers in time across up to 120 minutes of input.","keywords":["large audio language models","audio question answering","long-form speech processing","temporal grounding","inter-timing markers","time-aware audio LLM","synthetic supervision"],"falsifier":"Train an identical architecture without any inter-timing markers on the same synthetic data and measure long-form (20–40 min and 120 min) temporal-grounding mIoU; if it matches or exceeds the marker-equipped model, the claim that anchors are essential is false.","tokens_in":10863,"feed_emoji":"⏱️","tokens_out":625,"duration_ms":7206,"temperature":0.7,"pith_summary":"Long audio such as meetings and podcasts is only useful for interactive question-answering if the system can say not only what happened but when. This paper shows that ordinary audio language models fail at that task once recordings stretch beyond a few minutes: they invent unparseable timestamps or lose the timeline entirely. The authors fix the problem by inserting explicit time markers at regular intervals into the stream of continuous audio tokens, then training the model on large-scale synthetic question-answer pairs that force it to use those markers. The resulting model keeps high temporal accuracy from short clips out to two-hour inputs, and it can also produce time-anchored fragment descriptions and multi-part summaries. Ablations establish that both the markers themselves and a mixture of training lengths are necessary; without either ingredient, long-form grounding collapses.","feed_headline":"Time markers keep audio LLMs accurate across 120-minute recordings","feed_subtitle":"Without periodic anchors, long-form temporal grounding collapses; mixture training is also required.","key_machinery":"Inter-timings: periodic time markers (plain-text hh:mm:ss or dedicated tokens) inserted every N seconds into the continuous audio-token stream, acting as temporal anchors that the model can read and emit.","core_discovery":"Periodic temporal anchors interleaved with continuous audio tokens are essential for stable long-form grounding. With markers every 60 seconds the model reaches 53.8 mIoU on 20–40-minute temporal grounding (rising to 65.2 with 7-second markers), while removing the anchors drops the same metric to 14.2 and sharply degrades timed descriptions and summaries. Training on a mixture of short and long recordings is likewise required for length generalization up to 120 minutes; single-duration regimes do not transfer.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Periodic time markers enable audio LLM grounding over 120 minutes","Without anchors long-form audio temporal grounding collapses to 14.2 mIoU","Time markers every 60s lift 20–40-min audio grounding to 53.8 mIoU","Interleaved time tokens plus mixed durations train 2-hour audio LLMs","Mixture training needed for audio LLMs to generalize to 120-min inputs"],"cache_read_input_tokens":128,"weakest_assumption_plain":"Synthetic question-answer pairs generated by a text-only language model from automatically aligned transcripts, after slicing and a consistency verifier, are accurate enough for an audio model to learn genuine temporal localization rather than just matching transcript patterns.","fun_headline_variants_meta":{"raw":{"variants":["Periodic time markers enable audio LLM grounding over 120 minutes","Without anchors long-form audio temporal grounding collapses to 14.2 mIoU","Time markers every 60s lift 20–40-min audio grounding to 53.8 mIoU","Interleaved time tokens plus mixed durations train 2-hour audio LLMs","Mixture training needed for audio LLMs to generalize to 120-min inputs"]},"model":"grok-4.5","effort":"low","cost_usd":0.004112,"raw_usage":{"total_tokens":1220,"prompt_tokens":706,"num_sources_used":0,"completion_tokens":90,"cost_in_usd_ticks":41120000,"prompt_tokens_details":{"text_tokens":706,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":424,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":706,"tokens_out":90,"duration_ms":5113,"temperature":1.0,"reasoning_tokens":424,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T12:06:18.624846+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train an identical architecture without any inter-timing markers on the same synthetic data and measure long-form (20–40 min and 120 min) temporal-grounding mIoU; if it matches or exceeds the marker-equipped model, the claim that anchors are essential is false.","supporting_citations":[],"review_version":1}