{"id":"c29880b4-7c0b-43f4-b903-a08c4adafc29","arxiv_id":"2512.01208","paper_version":5,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A small complex-valued spectral model (PRISM) and a hybrid Wave-Particle Transformer are claimed to show that phase-based interference is a sufficient reasoning primitive, with a 4.94 vs 5.28 perplexity win on WikiText-103 that is confounded by asymmetric hyperparameters.","lead":"This paper introduces PRISM, a complex-valued language model that encodes information in phase angles and mixes tokens with Fourier transforms, and reports a hybrid architecture that slightly beats a Transformer on a masked-language benchmark. The mechanistic claim that phase-based 'subtractive interference' drives the gains is not supported by the experiments as reported.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"WPT-vs-Transformer gain is confounded by architecture and regularization; phase mechanism is not isolated.","rationale":"The reader identifies the same load-bearing weakness: the WPT-vs-Transformer comparison does not isolate phase coding. The concrete test would directly control for the most obvious confounds by substituting a real-valued spectral stream into the WPT topology. If the gap disappears, the central claim collapses. Even if the gap persists, the paper still owes the promised phase-disruption ablation to show that the phase component specifically carries task-relevant information. The reader's REJECT verdict is therefore appropriate; no change is needed.","tokens_in":14822,"tokens_out":3988,"duration_ms":44551,"concrete_test":"Train a matched control on WikiText-103 MLM: take the WPT topology but replace the complex PRISM relational stream with a real-valued FFT/FNet stream of identical parameter count, keeping RMSNorm, skip connections, refiner, and weight decay 0.0. If this phase-ablated control achieves PPL within 0.1 of WPT's 4.94 (or beats it), the reported gain is due to the extra architecture/regularization rather than subtractive phase interference. Additionally, if feasible, run the promised phase-scrambling ablation (randomize phases of the complex stream while preserving magnitudes) and compare against a magnitude-scrambling control.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central existence proof rests on Table 4: WPT 4.94 PPL vs Transformer 5.28. But §6.1 sets weight decay to 0.0 for PRISM Hybrid while baselines use 0.01; §6.2 gives WPT RMSNorm (vs LayerNorm), a dual-stream topology, skip connections, and a 1-layer attentive refiner absent from the Transformer baseline. These are not phase-specific changes. The abstract additionally promises an interventional phase-disruption ablation, but no such experiment appears in the manuscript; the phase-locking evidence (Tables 2, 5) is correlational, uses N≈120–133 pairs, and R can reflect ordinary embedding magnitude/similarity. The §5.4 L=1/L=2 'spectral starvation' claim also lacks a rate-coding baseline at those lengths, so repetition collapse may simply be insufficient context. Because the only quantitative superiority claim is the WPT gap, and that gap is not attributable to phase under the current design, the central mechanistic claim is unsubstantiated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces PRISM, a complex-valued architecture that imposes unit-norm phasors and gated harmonic convolutions, and claims that phase-based spectral interference can serve as a sufficient computational primitive for semantic reasoning. It reports phase-locking statistics, a qualitative spectral-density threshold, and a hybrid Wave-Particle Transformer (WPT) that achieves lower perplexity than a Transformer baseline. The paper concludes that subtractive interference is an algorithmic existence proof for semantic reasoning and proposes a blueprint for optical-digital co-processors.","tokens_in":15136,"tokens_out":6440,"duration_ms":67864,"significance":"If substantiated, the proposal would be a novel contribution to mechanistic interpretability and efficient sequence modeling. The paper has useful ingredients: it explicitly constrains magnitude to isolate phase, compares against an FNet spectral baseline, defines physical metrics, and releases code. However, the current experiments do not isolate the phase mechanism. The key WPT comparison is confounded by architectural and regularization differences, the abstract-promised interventional phase-disruption ablation is absent, and the iso-energetic gain evidence is partly circular. The significance therefore remains speculative until these controls are provided.","major_comments":[{"comment":"The central quantitative claim is the WPT improvement over the Transformer (4.94 vs 5.28 PPL). This comparison is not attributable to phase coding: §6.1 sets weight decay to 0.0 for PRISM Hybrid but 0.01 for baselines; §6.2 gives WPT RMSNorm instead of LayerNorm, a skip connection, a dual-stream topology, and a 1-layer attentive refiner, all absent from the Transformer baseline. In addition, the pure PRISM Hybrid is worse than the Transformer (6.06 vs 5.28 PPL), which undercuts the 'sufficient primitive' language in §8.2. A matched control with identical architecture and regularization except for the phase mechanism is required.","section":"§6.1–6.2, Table 4"},{"comment":"The abstract promises 'interventional ablations' showing that preserving phase largely maintains performance while disrupting phase causes severe degradation. No such experiment appears in the manuscript. §5.2 compares ambiguous and unambiguous tokens but does not perturb phase; §7 is correlational. Without an explicit phase-disruption intervention (e.g., randomizing angles while keeping magnitudes), the causal role of phase is not tested.","section":"Abstract; §5–§7"},{"comment":"The reported iso-energetic gain g≈1.0 is partly a consequence of the Phase-Preserving Layer Normalization in Eq. (8), which standardizes magnitudes, together with the enforced unit-norm constraint. Using g≈1.0 as empirical validation of phase coding is therefore circular. The metric should be computed on pre-normalization activations, or the claim should be removed.","section":"§3.5, Eq. (8); §5.2"},{"comment":"The claim that L=1/L=2 repetition collapse is due to 'spectral starvation' is not tested against a rate-coding baseline. A Transformer or FNet model at the same short lengths would also be expected to fail or repeat; without this control, the result is consistent with a trivial lack of context. No error bars or seeds are reported for this experiment, and the description is qualitative.","section":"§5.4"},{"comment":"The phase-locking differences are small (e.g., mean R 0.198 vs 0.072), computed on N=122–133 pairs, with no error bars, significance tests, or multiple seeds. Because Eq. (13) weights by |z_a,k||z_b,k|, the metric can be driven by ordinary magnitude/similarity correlations rather than phase. A control that randomizes phase while preserving magnitude is needed to establish that the coherence is phase-specific.","section":"§5.1, Eq. (13); Tables 2, 5"}],"minor_comments":[{"comment":"The text reports antonym mean R=0.117, while Table 2 reports 0.133. These values should be reconciled.","section":"§5.1 vs Table 2"},{"comment":"The text says FNet exhibits 'R≈0.61, Table 5', but Table 5 lists FNet synonym R=0.49, antonym R=0.61, random R=0.12. Please clarify which category is being quoted.","section":"§7 vs Table 5"},{"comment":"The notation ω_k=1/10000^{k/d} is described as a spectrum of geometric frequencies; please define the domain of k and the resulting units to avoid confusion with the positional phase ωt.","section":"§3.1, Eq. (1)"},{"comment":"The heading says 'Appendix H' while the section letter is G. The numbering should be corrected.","section":"Appendix G"},{"comment":"PRISM (U) has 128.4M parameters while the others are 69–73M; please clarify whether this is intentional and why it is included in the fairness discussion.","section":"§4.2, Table 1"}],"recommendation":"reject","confidential_remarks":"This is a work-in-progress preprint with a substantial gap between the abstract's causal claims and the experiments actually reported. The central existence proof is not supported because the main comparison is confounded and the promised interventional ablation is missing. I could see a resubmission after the authors run properly controlled experiments, but the current version does not meet the standard for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a well-organized exploration of complex-valued phase coding, and I want to give it credit where it's due. The specific architecture family (PRISM, HSSM, WPT) and the WikiText-103 numbers at 33M scale are new. They release code, they use FNet as a spectral control, they try to match parameter counts, and they are upfront that this is a mechanistic probe rather than a foundation-model race. The phase-locking measurements, synonyms > antonyms > random, are a reasonable sanity check even if small and possibly driven by embedding similarity. The rate-versus-phase distinction is worth taking seriously.\n\nBut the load-bearing claim does not hold up. The main result in Table 4 compares WPT against a Transformer baseline while changing weight decay (0.0 for PRISM variants, 0.01 for baselines), normalization (RMSNorm vs LayerNorm), topology (dual-stream, skip connections), and adding a 1-layer attention refiner. Those are not phase-specific changes, so the perplexity gap cannot be attributed to phase coding. The abstract promises an interventional phase-disruption ablation, but I could not find it in the body. That is a real omission, not just an overstatement. The L=1/L=2 \"spectral starvation\" experiment also lacks a rate-coding baseline at those lengths; repetition collapse could simply be insufficient context.\n\nSome smaller issues reinforce the pattern. The iso-energetic gain near 1.0 is partly a consequence of the Phase-Preserving Layer Normalization, which standardizes magnitudes, so it is not independent evidence that the model encodes meaning in phase. The R values in Table 2 are small (0.07–0.20) and come from about 120 pairs, without controlling for magnitude or frequency. And setting weight decay to zero for PRISM is itself a regularization difference that could affect generalization independently of phase.\n\nThese are not minor quibbles; the mechanistic conclusion rests on the confounded comparison. Still, the paper is coherent on its own terms and engages honestly with the literature. It deserves a serious referee, because the question matters and the authors have done real experimental work. The referee should reject the current version but ask for an actual phase-disruption ablation, matched baselines, and control experiments at L=1/L=2. If those come through, the story could be compelling. As it stands, I would not cite it as evidence for phase-based reasoning.","headline":"A clearly written complex-valued architecture paper with a promising idea, but the central claim that phase interference drives the reported gains is not supported by the current experiments.","tokens_in":15607,"tokens_out":1942,"would_cite":false,"duration_ms":21926,"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":"This paper argues that the phase angle of a complex-valued representation can carry semantic reasoning, and that subtractive interference is a sufficient primitive for language modeling at the evaluated scale.","keywords":["phase coding","subtractive interference","complex-valued neural networks","semantic phase locking","spectral reasoning","hybrid architecture","language modeling","mechanistic interpretability"],"falsifier":"Train a real-valued control that matches every non-phase ingredient of the Wave-Particle Transformer (skip connection, same normalization, dual-stream layout, attention refiner, weight decay 0.0) with magnitudes only and compare perplexity; or ablate the trained model by randomizing phase angles while preserving magnitudes. If the real-valued control matches 4.94 PPL, or if phase randomization leaves performance largely intact, the phase-mechanism claim is falsified.","tokens_in":1399,"feed_emoji":"🌊","tokens_out":1938,"duration_ms":57797,"temperature":0.7,"pith_summary":"The authors aim to separate two ways a neural network can encode meaning: by signal strength (magnitude, as in standard Transformers) or by signal direction (phase, as in waves). They build a complex-valued encoder, PRISM, that forces meaning into phase by requiring unit magnitude and using harmonic convolutions instead of attention. Their hybrid Wave-Particle Transformer mixes a small attention stream with a phase-based spectral stream and reaches 4.94 perplexity on a masked-language benchmark versus 5.28 for a standard Transformer, using 18% fewer core reasoning parameters. Mechanistic probes show that synonyms and antonyms become phase-aligned while random words do not, and that ambiguity resolution happens through a sharp phase rotation with near-constant magnitude. If the claim holds, it would establish that reasoning can be done by erasing unwanted signal through interference rather than only by amplifying wanted signal.","feed_headline":"Encoding words as waves beats standard attention","feed_subtitle":"A phase-based model hits 4.94 perplexity vs 5.28 for a Transformer, using 18% fewer core reasoning parameters.","key_machinery":"The key machinery is the Phase-Rotating Interference Spectral Model (PRISM), a complex-valued sequence encoder that treats tokens as phasors z = r·e^(iθ) with semantic identity in the angle θ. A strict unit-norm constraint (|z| = 1) removes magnitude as a carrier of importance. Attention is replaced by a Gated Harmonic Convolution: the sequence is Fourier-transformed, multiplied by a learnable global filter, transformed back, and gated separately on real and imaginary parts so the phase angle can be steered. The Wave-Particle Transformer extends this by running a small attention-based 'sensory' stream in parallel with a phase-based 'relational' stream and fusing them with a one-layer attenti","core_discovery":"The central discovery is that semantic relationships can be encoded geometrically as phase alignment in a fixed-magnitude complex space: synonyms lock phase, antonyms align too, and random pairs sit at a noise floor. The model resolves polysemy by applying a heavy-tailed phase rotation at a specific layer while keeping gain near 1.0, which the authors interpret as ruling out magnitude-based 'shouting'. The headline result is that the Wave-Particle Transformer, which combines a lightweight attention stream with a phase-based spectral stream, achieves 4.94 perplexity on WikiText-103 masked reconstruction versus 5.28 for a standard Transformer, with 18% fewer core reasoning parameters. The auth","pith_inferences":["Our inference: if phase coherence is genuinely the carrier of semantic relatedness, interpretability tools could track phase alignment rather than attention weights or activation magnitudes, and phase coherence could become a diagnostic for semantic closeness in complex-valued models.","Our inference: the 'spectral starvation' account of short-input collapse is testable by running a rate-based recurrent baseline on isolated tokens; if that baseline also collapses, the phenomenon may be about insufficient context rather than specifically about phase.","Our inference: the architecture points toward a concrete hardware roadmap—a passive optical unit for O(N log N) interference and a small digital unit for attention—but the paper itself does not demonstrate such a device, only the algorithmic blueprint.","Our inference: the phase-locking of antonyms suggests that the model groups concepts by shared topic rather than by polarity, which, if general, would imply that phase encodes topical relatedness and that logical opposition is represented by an offset within a common phase band."],"forward_implications":["If phase can carry semantic reasoning, a model can suppress noise by destructive interference instead of amplifying signal, suggesting that reasoning requires the capacity to erase, not just accumulate.","Hybrid architectures that decouple intensity (magnitude) from direction (phase) can match or beat a standard Transformer with fewer active reasoning parameters, implying that not every reasoning operation needs quadratic attention.","Sequence length matters not only as context but as spectral resolution: longer sequences provide denser interference fields, so the model's performance should continue to benefit from long inputs in ways that rate-only models may not.","A single lightweight attention layer at the output can serve as a readout for a phase-based encoder, suggesting that deep attention stacks can be partially replaced by passive spectral mixing.","Purely additive spectral mixing is insufficient for contextual reasoning at long sequence lengths; a phase-based subtractive component appears necessary to prevent collapse."],"fun_headline_variants":["Phase beats magnitude in wave-based language model","Wave-particle model: phase encoding cuts perplexity","Semantic phase locking: wave model trims params, beats transformer","Language as waves: phase interference wins with fewer params","Wave model encodes meaning in phase, outperforms attention"],"cache_read_input_tokens":17024,"weakest_assumption_plain":"The central claim rests on the assumption that the perplexity advantage of the hybrid model comes from phase interference itself, rather than from the extra architectural and regularization differences—the skip connection, the different normalization, the dual-stream topology, the added attention refiner, and the use of weight decay 0.0 for the phase model versus 0.01 for the baselines.","fun_headline_variants_meta":{"raw":{"variants":["Phase beats magnitude in wave-based language model","Wave-particle model: phase encoding cuts perplexity","Semantic phase locking: wave model trims params, beats transformer","Language as waves: phase interference wins with fewer params","Wave model encodes meaning in phase, outperforms attention"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000218,"raw_usage":{"total_tokens":1252,"prompt_tokens":696,"completion_tokens":556,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":440,"completion_tokens_details":{"reasoning_tokens":480}},"tokens_in":440,"tokens_out":556,"duration_ms":5981,"temperature":1.0,"reasoning_tokens":480,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T19:16:34.442791+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a real-valued control that matches every non-phase ingredient of the Wave-Particle Transformer (skip connection, same normalization, dual-stream layout, attention refiner, weight decay 0.0) with magnitudes only and compare perplexity; or ablate the trained model by randomizing phase angles while preserving magnitudes. If the real-valued control matches 4.94 PPL, or if phase randomization leaves performance largely intact, the phase-mechanism claim is falsified.","supporting_citations":[],"review_version":1}