{"id":"e8c10cc6-cbb9-469d-be5c-32a7b48d2cf4","arxiv_id":"2607.08800","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Uncorrelated dither noise lets dual frozen BEATs encoders preserve inter-channel amplitude differences across LLM normalizers, yielding up to 97% left/center/right accuracy and zero-shot spatial generalization.","lead":"Standard audio LLMs ignore stereo space because mono downmixing and internal normalizers erase left-right differences. Dual-BEATs routes channels separately and adds uncorrelated dither noise so the model can localize sound left/center/right at high accuracy, even zero-shot.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The load-bearing claim of native stereo understanding rests on synthetic ICLD-only panning that may not survive real stereo/binaural geometry.","rationale":"The reader correctly isolates the synthetic amplitude-panning assumption as the weakest load-bearing point. All headline numbers (Table 1, Figure 2) and the causal story about the normalization bottleneck are obtained under pure ICLD. The undithered collapse and seed-isolation controls make the dither intervention credible inside that regime, but they do not license the leap to “natively capable of generalized stereo audio understanding.” Because the paper already lists the limitation honestly, the appropriate stance remains CONDITIONAL rather than REJECT; the concrete HRTF transfer test would decide whether the claim can be strengthened or must be narrowed. No stronger internal inconsistency appears in the reported experiments.","tokens_in":14805,"tokens_out":548,"duration_ms":15499,"concrete_test":"Take a held-out subset of the same AudioSet clips, render them with a standard HRTF (e.g., CIPIC or SADIE) at discrete azimuths that map to Left/Center/Right, keep the identical Dual-BEATs + DA=0.05 pipeline and LoRA adapters trained only on amplitude-panned data, and measure ternary directional accuracy. If accuracy falls near chance or loses the flat zero-shot curve of Figure 2d, the synthetic-ICLD proxy is insufficient for the generalized claim.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim (97%+ accuracy at PA=0.50 + continuous zero-shot generalization proving that standard multimodal models are natively capable of generalized stereo audio understanding once geometry is protected from normalization) is demonstrated exclusively under deterministic single-source amplitude panning of mono AudioSet clips (§3.3, A.3). This produces pure Inter-channel Level Differences (ICLD) with no HRTF filtering, no interaural time/phase differences, no multi-source overlap, and no room acoustics. The Dual-BEATs + uncorrelated dither mechanism is therefore only shown to preserve and allow the LLM to read a simple left/right gain ratio after BEATs encoding and projector normalization. Section 5 correctly flags this as a limitation, yet the abstract, introduction, and conclusion still generalize to “true spatial abstraction” and “generalized stereo audio understanding.” If real stereo or binaural signals present additional phase/spectral structure that interacts differently with BEATs’ spectro-temporal features or with the same normalizers, the dither bridge may not transfer and the “native capability” claim would not hold beyond the synthetic proxy.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes Dual-BEATs: left and right stereo channels are routed independently through two frozen identical BEATs encoders, feature-concatenated, and projected into an LLM (Gemma-3-1B or OLMo-3-7B) via QLoRA. The authors identify internal variance normalizers (LayerNorm/RMSNorm) as a bottleneck that erases subtle inter-channel level differences (ICLD), and introduce static uncorrelated Gaussian dither (DA=0.05) prior to encoding as a Stochastic Resonance bridge that preserves spatial geometry. On ternary Left/Center/Right classification of amplitude-panned AudioSet clips, dithered OLMo-3-7B reaches ~97% accuracy at PA=0.50 and shows continuous zero-shot generalization across unseen panning amplitudes (Fig. 2), while undithered models collapse to chance near center pan (Table 1). A modest semantic F1 tax is reported (Fig. 3). The authors conclude that standard multimodal models are natively capable of generalized stereo understanding once geometry is protected from normalization.","tokens_in":15125,"tokens_out":1348,"duration_ms":16883,"significance":"If the mechanism transfers beyond the synthetic proxy, the work offers a modular, encoder-agnostic path to stereo perception in audio LLMs without custom geometry-aware encoders or room simulation, which would be practically valuable for accessibility and scene analysis. Strengths include clean dithered-vs-undithered contrasts (Table 1), zero-shot heatmaps (Fig. 2), explicit seed isolation and uncorrelated-noise controls against shortcut learning (Appendix A.3–A.4), and transparent documentation of the semantic tax and limitations. The normalization-bottleneck diagnosis and signal-level dither intervention are concrete and falsifiable within the ICLD setting. The contribution is therefore of genuine interest to the audio-LLM community, provided claims are scoped to what the experiments actually demonstrate.","major_comments":[{"comment":"Abstract, §1, and §6 claim “true spatial abstraction,” “generalized stereo audio understanding,” and that standard models are “natively capable” of stereo perception. All positive results (Table 1, Fig. 2) are obtained exclusively under deterministic single-source amplitude panning of mono AudioSet clips (ICLD only; §3.3, A.3). There is no HRTF, ITD/phase, multi-source overlap, or real stereo/binaural evaluation. §5 correctly flags this as a limitation, but the framing in the abstract and conclusion overstates the evidence. Either (i) temper the central claim to “ICLD-based directional perception under amplitude panning” or (ii) add at least one transfer experiment on real stereo mixes or simple binaural renderings. Without one of these, the load-bearing “native generalized stereo” claim is not supported.","section":"Abstract, §1, §6 vs §5 / A.3"},{"comment":"The dither amplitude is fixed at DA=0.05 with no systematic ablation of the noise floor (only noted as future work in §5). Because the rescue effect is attributed to establishing a macro-variance floor that survives normalization, the result is sensitive to this free parameter. A minimal DA sweep (e.g., 0.01–0.10) on at least one backbone, reporting both directional accuracy and semantic F1, is needed to show that the Stochastic Resonance bridge is robust rather than a single-point hyperparameter success.","section":"§3.2, §4.1, §5"},{"comment":"There is no head-to-head comparison against a specialized spatial encoder (e.g., Spatial-AST / BAT or a simple dual-channel baseline with phase features) on the same ternary task and AudioSet subset. Without this, it is hard to judge whether Dual-BEATs + dither is competitive or merely “works when specialized methods are unavailable.” A single controlled baseline would strengthen the accessibility claim in the introduction.","section":"§2, §4"}],"minor_comments":[{"comment":"Table 1: Gemma-3-1B shows non-monotonic and sometimes degraded accuracy when dither is on at PA=0.00 (e.g., Direction-First 66.8% → 47.7%). A short discussion of noise-to-signal sensitivity for smaller models would help readers interpret the architecture’s capacity dependence.","section":"Table 1, §4.1"},{"comment":"Figure 3: the mono BEATs baseline (40.6%) and “Music; Speech” prior (28.73%) are useful anchors; state explicitly in the caption whether the mono baseline was trained under the same QLoRA recipe or is an off-the-shelf reference.","section":"Figure 3"},{"comment":"Notation: P A and D A are written with spaces inconsistently (PA / P A, DA / D A). Standardize to PA and DA throughout.","section":"Throughout"},{"comment":"Appendix A.1: the frozen March 2024 AudioSet snapshot is good for reproducibility; consider releasing the exact YouTube ID list or a DOI-backed artifact so others can reconstruct the 18,373 / 17,148 splits.","section":"Appendix A.1"},{"comment":"Related work: “The World is Not Mono” [18] is cited; a one-sentence contrast with Dual-BEATs (dual semantic encoders + dither vs. heterogeneous semantic+spatial encoders) would clarify positioning.","section":"§2"}],"recommendation":"major_revision","confidential_remarks":"The technical core (normalization bottleneck + uncorrelated dither) is interesting and the controls against shortcut learning are better than average. The main risk is overclaim relative to a pure ICLD synthetic proxy; if the authors scope the language and add a small real-stereo or DA-ablation experiment, this could be a solid contribution. Fit for a cs.SD / audio-ML venue is good; less so for a general ML venue without broader validation."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful result here is simple and well-controlled: route left and right through two frozen BEATs, inject uncorrelated static dither (DA=0.05), and ordinary LLMs (Gemma-1B, OLMo-7B) suddenly keep left/center/right accuracy high even at PA=0.5 and transfer continuously across unseen panning amplitudes. Undithered runs collapse to chance at the center, which matches their normalization-bottleneck story. Seed isolation and the uncorrelated-noise control against the identity shortcut are done properly. Tables 1 and the heatmaps in Fig. 2 are the real contribution; the semantic-tax curves in Fig. 3 keep the cost honest.\n\nWhat is new is the modular intervention itself—dual identical semantic encoders plus a noise floor framed as a Stochastic Resonance bridge past LayerNorm/RMSNorm—rather than another geometry-aware spatial encoder. Related work already has specialized spatial modules and hybrid setups; this paper shows you can get usable ternary localization without them, at least under pure ICLD panning of AudioSet mono clips. The math is light (just additive independent Gaussians), the data pipeline is reproducible from the appendix, and the citation pattern is fair.\n\nThe soft spot is exactly the one the stress-test flags, and the authors already list it in §5: everything is deterministic single-source amplitude panning. No HRTF, no ITD/phase, no multi-source, no rooms. So the 97% number and the “native stereo understanding” language in the abstract and conclusion are stronger than the evidence. The mechanism is shown to protect a left/right gain ratio, not full stereo geometry. Fixed DA, no error bars, and no public code are minor relative to that. The central causal claim (dither vs. undithered on the same stack) still holds inside the synthetic regime they actually tested.\n\nThis is for people building audio LLMs or accessibility/spatial-QA tools who want a cheap modular fix rather than a new encoder. It deserves a serious referee; the experiments are sharp enough that the overclaim can be dialed back in revision. I would engage with it and cite the dual-encoder + dither result when the synthetic-to-real gap is acknowledged.","headline":"Clean dual-encoder + uncorrelated dither result that really does rescue center-pan localization and zero-shot panning transfer; the overclaim is only that this equals generalized stereo understanding.","tokens_in":15696,"tokens_out":564,"would_cite":true,"duration_ms":10161,"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":"Uncorrelated dither lets ordinary dual audio encoders give multimodal LLMs native stereo localization without specialized spatial modules.","keywords":["Dual-BEATs","stereo audio perception","dithering","normalization bottleneck","zero-shot spatial localization","multimodal LLMs","inter-channel level difference","Stochastic Resonance"],"falsifier":"Train and evaluate the same Dual-BEATs + dither pipeline on true binaural recordings that include HRTFs and multi-source overlap; if directional accuracy collapses to chance while the mono-panned condition remains high, the central claim that the method unlocks general stereo perception fails.","tokens_in":15681,"feed_emoji":"🎧","tokens_out":819,"duration_ms":7567,"temperature":0.7,"pith_summary":"Multimodal audio LLMs understand what a sound is, but they are built on mono pipelines and therefore cannot hear where it is. The paper shows that this spatial blindness is not inevitable: simply routing left and right channels through two identical frozen semantic encoders fails only because internal normalization layers erase the tiny inter-channel amplitude differences that encode panning. Adding a static, uncorrelated dither noise floor before encoding creates a macro-variance that carries those differences through the normalizers, so ordinary cross-attention can read left-right geometry. On a ternary Left/Center/Right task the dithered models reach roughly 97 percent accuracy even at subtle 0.5 panning and generalize zero-shot to panning levels never seen in training. The practical claim is that, with this acoustic regularization, standard off-the-shelf multimodal models already possess generalized stereo understanding and do not need custom geometry-aware encoders.","feed_headline":"Dither turns ordinary dual encoders into stereo ears for audio LLMs","feed_subtitle":"Uncorrelated noise floor carries panning cues past normalizers; 97% accuracy at subtle 0.5 panning, zero-shot.","key_machinery":"Dual-BEATs with uncorrelated dither: left and right waveforms each receive independent Gaussian noise scaled by a fixed dithering amplitude (DA=0.05), are passed through identical frozen BEATs encoders, feature-concatenated, and projected into the LLM; the uncorrelated noise floor \"smuggles\" spatial geometry across LayerNorm/RMSNorm.","core_discovery":"Standard multimodal audio LLMs can perform continuous stereo localization once left and right channels are independently encoded and an uncorrelated dither floor is injected; the dither establishes a macro-variance that prevents internal normalizers from equalizing inter-channel amplitude differences, unlocking up to 97 percent ternary accuracy at subtle panning and robust zero-shot transfer to unseen spatial configurations.","pith_inferences":["The same variance-floor idea may protect other continuous multimodal differentials that normalizers currently erase, such as mild spectral or temporal asymmetries.","If multi-source or true binaural tests succeed, the method could become a default preprocessing step for any audio LLM rather than a specialized research module.","Adversarial control of the dither seed is an immediate security surface once the technique moves to edge devices.","A lower DA schedule might shrink the observed semantic tax while still spanning the normalization threshold."],"forward_implications":["Stereo localization becomes a lightweight adapter step rather than a full re-design of the audio encoder.","Existing mono-pretrained multimodal models can be given spatial awareness without discarding their semantic weights.","Zero-shot generalization across continuous panning amplitudes becomes the expected default once the noise floor is present.","The same dual-encoder-plus-dither pattern can be tested on larger backbones and denser token rates without new spatial pre-training.","Accessibility tools that need to name left-right events can reuse ordinary open multimodal stacks."],"fun_headline_variants":["Dual-BEATs: dual encoders plus dither unlock stereo in audio LLMs","Dither noise floor carries panning past normalizers for stereo LLMs","Uncorrelated dither smuggles spatial cues through dual-BEATs encoders","Independent L/R encoding and static dither enable zero-shot stereo","Dithered dual encoders hit 97% ternary stereo accuracy zero-shot"],"cache_read_input_tokens":2432,"weakest_assumption_plain":"Deterministic single-source amplitude panning of mono clips is treated as a faithful enough stand-in for the spatial cues that real stereo or binaural audio would present to the same encoders and normalizers.","fun_headline_variants_meta":{"raw":{"variants":["Dual-BEATs: dual encoders plus dither unlock stereo in audio LLMs","Dither noise floor carries panning past normalizers for stereo LLMs","Uncorrelated dither smuggles spatial cues through dual-BEATs encoders","Independent L/R encoding and static dither enable zero-shot stereo","Dithered dual encoders hit 97% ternary stereo accuracy zero-shot"]},"model":"grok-4.5","effort":"low","cost_usd":0.006908,"raw_usage":{"total_tokens":1740,"prompt_tokens":787,"num_sources_used":0,"completion_tokens":85,"cost_in_usd_ticks":69080000,"prompt_tokens_details":{"text_tokens":787,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":868,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":787,"tokens_out":85,"duration_ms":6193,"temperature":1.0,"reasoning_tokens":868,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T06:39:42.589626+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train and evaluate the same Dual-BEATs + dither pipeline on true binaural recordings that include HRTFs and multi-source overlap; if directional accuracy collapses to chance while the mono-panned condition remains high, the central claim that the method unlocks general stereo perception fails.","supporting_citations":[],"review_version":1}