{"id":"4e820ead-4ddb-4c3c-bd2d-b160fdeeaf01","arxiv_id":"2608.13365","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":1,"one_line_summary":"Under dynamic W4A4KV4 quantisation, a head-shared RoPE-aligned pairwise rotation that exactly minimizes a pooled variance surrogate still yields higher perplexity than full-head Hadamard mixing in all evaluated comparisons.","lead":"This paper tests whether a rotation that respects RoPE's built-in two-channel structure can improve 4-bit quantization of attention in large language models, and finds it does not in the tested setup. It derives the optimal such rotation for a simplified variance objective, then shows that optimality for that objective fails to reduce quantization error because the actual quantizer sets its scale from a different statistic.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: the negative result is explicitly scoped to the head-shared pairwise subfamily, and the paper's claims stay inside that boundary.","rationale":"The reader's weakest_assumption correctly identifies the head-shared subfamily as the tested parameterisation and notes that head-dependent angles are untested. I agree that this is the most fragile premise if one wanted to generalise beyond the paper's explicit scope. However, the paper consistently and explicitly confines its empirical and theoretical claims to the head-shared configuration: the abstract says 'the evaluated head-shared pairwise configuration,' Section 3 states that the implementation imposes the head-shared constraint, and Limitations item 1 says 'Head-dependent angles φ_{h,k} and transforms that mix across heads are untested.' The central claim therefore does not depend on head-dependent angles behaving the same way. The algebraic results are carefully scoped to the pooled-covariance, position-averaged surrogate, with the position-stationarity caveat stated before Theorem 3 and the verification caveat stated after it. The main empirical ordering is consistent across checkpoints and contexts, with the strongest Llama-3.2-3B result supported by paired intervals. The long-context ranges without confidence intervals are a reporting limitation, but they are descriptive robustness evidence rather than the load-bearing part of the argument. I therefore find no significant objection that would change the reader's conditional verdict.","tokens_in":41402,"tokens_out":9870,"duration_ms":110370,"concrete_test":"Run a head-dependent angle probe on Llama-3.2-3B: estimate per-head, per-frequency-pair angles from head-resolved calibration covariances (rather than pooling across heads), apply the resulting head-dependent pairwise rotation under the same dynamic W4A4KV4 protocol, and compare perplexity and K range against full-head Hadamard. If head-dependent pairwise rotations also fail to match full-head Hadamard, the head-shared boundary is immaterial to the conclusion; if they succeed, the paper's framing would need to be narrowed further.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is a scoped empirical negative result: the evaluated head-shared RoPE-aligned pairwise rotation does not improve dynamic W4A4KV4 accuracy relative to full-head Hadamard mixing. The theoretical component (Lemma 2, Theorem 3) is internally consistent, the surrogate optimum is numerically verified on Llama-3.2-3B, and the main paired perplexity comparisons support the stated conclusion. The most plausible broader concern is that head-dependent angles, which Lemma 2 explicitly permits but the implementation does not test, could behave differently and rescue pairwise rotations. However, the abstract, Section 5, and the Limitations all repeatedly restrict the empirical conclusion to the 'evaluated head-shared configuration' and explicitly list head-dependent angles and cross-head mixing as untested. The paper therefore does not overclaim the unverified generalisation. A secondary concern is that Table 4 and Table 20 report long-context point-estimate ranges without confidence intervals; this affects the strength of the cross-checkpoint robustness wording, but the primary Llama-3.2-3B result carries paired intervals and is sufficient for the central negative claim. Neither concern undermines the argument as scoped.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies the design of the online Q/K rotation R3 in dynamic W4A4KV4 post-training quantisation, contrasting rotations that respect the RoPE frequency-pair structure with full-head Hadamard mixing. The theoretical part characterises the single-head centraliser of RoPE as the product of independent per-pair rotations (Lemma 2, including the converse), derives a closed-form angle minimising the larger per-pair channel variance of a pooled-covariance, position-averaged surrogate (Theorem 3), and reports numerical attainment of that minimum on Llama-3.2-3B. The empirical part compares the head-shared pairwise configuration with the full-head Hadamard across four checkpoints and multiple context lengths: pairwise-only has higher perplexity in every short-context paired comparison; composing it with the Hadamard meets a pre-specified ±0.05 PPL interval criterion; K-only angle estimation improves pairwise-only without closing the gap; and a block-Hadamard support interpolation associates larger mixing support with lower K range, lower relative quantisation error, and smaller perplexity degradation.","tokens_in":41598,"tokens_out":12390,"duration_ms":133074,"significance":"If the results hold, the paper provides a clean, explicitly scoped negative benchmark: exact optimality for a natural RoPE-aligned surrogate does not translate into improved accuracy under the evaluated dynamic quantiser, and the discrepancy is plausibly tied to the quantiser's tokenwise range statistic and two-channel mixing support. The strengths are the clean algebra in Lemma 2 and Theorem 3, the numerical attainment check (maximum excess 6.90e-7 against a 5e-5 tolerance), the paired seed design for the primary short-context comparisons with confidence intervals, and the unusually careful boundary-drawing: the empirical conclusion is repeatedly restricted to the head-shared subfamily, with head-dependent angles and cross-head mixing explicitly listed as untested. The reproducibility apparatus, including portable drivers, manifest generators, and verification scripts, is a real asset. The principal limitation is that the negative result is for the head-shared subfamily; the paper does not overclaim beyond that.","major_comments":[],"minor_comments":[{"comment":"The long-context cross-model evidence in Table 20 is reported as point-estimate ranges without paired confidence intervals or standard deviations. Since the abstract and Section 7 state that pairwise-only increases perplexity at both short and long context lengths across four checkpoints, please either supply the paired interval information for these points or explicitly restrict the long-context claim to mean point estimates for the non-primary checkpoints.","section":"Section 7 / Table 20"},{"comment":"The function wrap[−π/4,π/4) is used in Theorem 3 but is never defined. Please define the residue-modulo-π/2 convention at first use; the proof in Appendix C uses it, but the body should stand alone.","section":"Section 3, Eq. (3)"},{"comment":"The band-k components of q and k are written q(k) and k(k), which collides with the band index k. Please rename the key vector, for example as κ^(k) or kbold^(k), for readability.","section":"Appendix C, proof of Proposition 4"},{"comment":"The verification that the implementation attains the analytic minimum reconstructs the surrogate from stored per-pair covariances; this is disclosed in the appendix, but a one-sentence statement in the main text saying that the verification uses reconstructed moments rather than raw activations would prevent over-reading of the attainment claim.","section":"Section 4 / Appendix A.5"},{"comment":"In the support interpolation, block sizes b>2 deliberately leave the RoPE-commuting family, so the monotone improvement in Table 3 conflates support size with loss of commutativity. The text's caveat is explicit, but the summary's 'as support increases' phrasing invites a causal reading; consider adding 'along the evaluated non-commuting block-Hadamard path' to the summary.","section":"Section 6.3 / Summary"}],"recommendation":"minor_revision","confidential_remarks":"I have no concerns about the citation pattern or novelty disclosure; FPTQuant is credited for the commuting family and the converse is clearly labelled. This is a well-scoped empirical paper with unusually careful limitations. My recommendation is minor revision based on presentation-level issues in the long-context reporting and some notation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nPunchline: this is a scoped negative result that holds up. The head-shared RoPE-aligned pairwise rotation does not improve dynamic W4A4KV4 perplexity over full-head Hadamard, and the authors do not oversell it.\n\nWhat's actually new: Lemma 2 gives the converse centralizer characterization for distinct frequencies, and Theorem 3 gives a closed-form angle for the head-shared pooled-covariance surrogate. The numerical attainment check (max excess 6.9e-7) is real evidence that they optimized exactly what they claim. But the more valuable content is the negative evaluation plus the diagnostics: the exact minimizer of a position-averaged second-moment surrogate fails to reduce the quantizer's tokenwise range, and mixing support matters. The block-Hadamard interpolation is a nice, controlled addition.\n\nWhy I trust it: paired within-seed comparisons, shared calibration and seeds, primary claims carry confidence intervals, and the limitations are unusually explicit. The head-shared parameterization is clearly distinguished from the head-dependent family, and the paper repeatedly restricts its conclusions to what was tested. No circularity: they never treat surrogate optimality as evidence of perplexity improvement. The math is clean.\n\nSoft spots, in proportion: Table 4 and Table 20 report long-context point-estimate ranges without intervals; minor, because the primary Llama-3.2-3B claims have paired intervals and are sufficient for the central negative result. The ±0.05-PPL equivalence margin is self-selected and not externally justified; they disclose it, but it makes the \"satisfies the criterion\" language weaker than it looks. The static control is transductive and exploratory, and they label it as such. None of these are load-bearing.\n\nWho this is for: anyone designing post-RoPE Q/K transforms or relying on surrogate objectives for quantization. It deserves a serious referee. I would send it to review and let the referees ask for the missing long-context intervals if they want them.","headline":"A scoped negative result that holds up: the head-shared RoPE-aligned pairwise rotation does not beat full-head Hadamard under dynamic W4A4KV4, and the paper is honest about that scope.","tokens_in":42160,"tokens_out":1265,"would_cite":true,"duration_ms":16106,"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":"RoPE-commuting rotations are exactly the per-frequency-pair rotations, yet the closed-form optimum of that family fails to improve 4-bit quantisation because its surrogate is misaligned with the quantiser's tokenwise range statistic.","keywords":["post-training quantisation","RoPE-aligned rotations","W4A4KV4","Hadamard rotation","KV cache quantisation","mixing support","covariance surrogate","perplexity evaluation"],"falsifier":"Re-run the exact W4A4KV4 protocol with head-dependent per-pair angles $\\phi_{h,k}$, which the paper's Lemma 2 allows: if any such configuration matches or beats the full-head Hadamard baseline on WikiText-2 perplexity, the negative result is an artefact of the head-shared constraint rather than of RoPE-aligned pairwise rotation itself. Alternatively, optimise the rotation against the quantiser's actual tokenwise group range statistic instead of the position-averaged second moment; if perplexity then reaches the Hadamard level, the surrogate mismatch — not the two-channel support — is the operative cause.","tokens_in":41142,"feed_emoji":"📉","tokens_out":17918,"duration_ms":146791,"temperature":0.7,"pith_summary":"Rotation-based post-training quantisation of LLMs usually applies an orthogonal transform across an entire attention head, while rotary position embeddings (RoPE) partition each head into two-dimensional frequency pairs; this paper asks whether a rotation that respects that pair structure can beat the usual full-head Hadamard mix. It proves the converse characterisation that, for pairwise-distinct frequencies, the only single-head orthogonal maps commuting with RoPE are independent rotations inside each frequency pair, derives the exact angle that minimises the larger channel variance of a position-averaged pooled-covariance surrogate, and verifies that its implementation attains that analytic minimum. The empirical answer is negative: across four checkpoints under dynamic W4A4KV4 quantisation, the head-shared pairwise rotation — even at its verified optimum — raises perplexity relative to the full-head Hadamard at every context length tested. Composing the pairwise rotation with the Hadamard stays within a selected ±0.05 perplexity equivalence margin, and estimating the angle from the K stream alone narrows but does not close the gap. The paper's explanation is that the rotation optimises a second-moment surrogate while the dynamic quantiser sets its step from a tokenwise group range, and the two-channel mixing support of the pairwise rotation cannot redistribute a peak the way a full head can.","feed_headline":"Even the exact optimum pair rotation loses to full-head mixing","feed_subtitle":"Under 4-bit Q/K quantisation, the rotation's variance surrogate never controls the tokenwise range the quantiser uses.","key_machinery":"The load-bearing objects are the single-head RoPE centraliser and the pooled-covariance surrogate it is paired with. Lemma 2 identifies $\\mathcal{B}_{d_h}=\\bigoplus_{k=1}^{K}G(\\phi_k)\\cong SO(2)^{K}$, with $K=d_h/2$, as the complete family of orthogonal maps commuting with RoPE when the frequencies are distinct, turning 'RoPE-aligned' into a sharply bounded design space whose freedom is exactly one angle per pair. On that space, the surrogate $\\tilde\\Sigma^{(k)}=\\tfrac{1}{L}\\sum_{m=0}^{L-1}R^{(k)}_m\\Sigma^{(k)}(R^{(k)}_m)^\\top$ position-averages the per-pair calibration covariance, and Theorem 3 solves its minimax channel-variance problem in closed form, giving the angle that equalises the two rotated diagonal entries; Proposition 4 then pins the Q/K application rule, since shared band angles are necessary and sufficient to preserve relative-position attention scores. The explanatory concept of the paper is 'mixing support' — the number $b$ of channels over which a peak can be redistributed, $b=2$ for the pairwise family versus $b=d_h$ for the Hadamard — which the block-Hadamard interpolation uses to show that range, quantisation error, and perplexity all fall as support grows.","core_discovery":"On the paper's own terms, the central discovery is a scoped negative result: exact optimality of a RoPE-aligned local objective does not translate into quantisation accuracy. For a single head with pairwise-distinct RoPE frequencies, the commuting orthogonal maps in $O(d_h)$ are exactly the band-aligned rotations $\\mathcal{B}_{d_h} \\cong SO(2)^{d_h/2}$ acting independently inside each frequency pair (Lemma 2), the converse of the per-pair construction the paper attributes to earlier work. For the implemented head-shared subfamily, Theorem 3 gives the closed-form equalising angle $\\phi_k^* = \\mathrm{wrap}_{[-\\pi/4,\\pi/4)}\\big[\\tfrac12\\operatorname{atan2}(\\tilde\\Sigma^{(k)}_{11}-\\tilde\\Sigma^{(k)}_{22},\\,2\\tilde\\Sigma^{(k)}_{12})\\big]$ that minimises $\\max_{j\\in\\{1,2\\}}\\big[G(\\phi_k)\\tilde\\Sigma^{(k)}G(\\phi_k)^\\top\\big]_{jj}$ down to $(\\sigma_1^2+\\sigma_2^2)/2$, and the implementation attains this minimum to within $6.9\\times10^{-7}$ against a tolerance of $5\\times10^{-5}$. Yet pairwise-only replacement of the full-head Hadamard increases perplexity on all four checkpoints at every short- and long-context point, and the verified surrogate optimum $\\phi_k^*$ has the largest pairwise-only gap on Llama-3.2-3B. The account offered is a double misalignment: the surrogate averages a per-pair second moment over positions, while dynamic INT4 sets $\\Delta_{t,g}=(u_{t,g}-\\ell_{t,g})/(2^B-1)$ from a tokenwise group range, and the pairwise transform's mixing support is only two channels, where the outlier-suppression bound is $1/\\sqrt{2}$ instead of roughly $1/\\sqrt{128}$ for a full head; a block-Hadamard interpolation from two channels to the full head monotonically reduces K range, relative quantisation error, and perplexity degradation.","pith_inferences":["Editorial inference: because the paper's own Lemma 2 permits head-dependent angles $\\phi_{h,k}$, and only the head-shared subfamily was tested, the negative conclusion may not extend to the full RoPE-commuting family; an evaluation with per-head pairwise angles would settle that boundary.","Editorial inference: the paper's mismatch analysis suggests a direct alternative design — choose the rotation to minimise the quantiser's actual tokenwise group range at full broadcast support rather than a position-averaged second moment — which is a testable construction the paper does not run.","Editorial inference: the exploratory static per-channel control reversed the perplexity ordering, so the negative result is tied to dynamic per-token quantisation and may not transfer to static scale rules or other bit widths.","Editorial inference: since post-RoPE Q/K maps need not commute with RoPE to preserve correctness (the same transform on both streams preserves inner products), the monotone support interpolation points toward head-wide or learned rotations as the more promising RoPE-aware direction, with commutativity available as a constraint only when it does not shrink mixing support."],"forward_implications":["A rotation for quantisation should be judged by whether its optimisation objective matches the quantiser's scale-setting statistic: commuting with RoPE is a structural constraint that, on its own, gives no accuracy benefit in the tested dynamic W4A4KV4 setting.","K-only estimation of the shared angle outperforms Q/K-pooled estimation on all four checkpoints, so the angle should be estimated from the stream that is actually quantised, though even this does not reach the full-head Hadamard baseline.","Mixing support bounds peak suppression: the post-rotation peak magnitude is at least $|x|/\\sqrt{b}$, giving $|x|/\\sqrt{2}$ at $b=2$ versus $|x|/\\sqrt{128}$ at a full head, and the evaluated support interpolation from $b=2$ to $b=d_h$ monotonically lowers K range, relative quantisation error, and perplexity degradation.","Exact attainability of a surrogate optimum certifies the surrogate, not the deployment: the verified optimum $\\phi_k^*$ produced the largest pairwise-only perplexity gap on Llama-3.2-3B.","Pairwise rotations remain usable as a layer on top of full-head mixing, since pairwise+Hadamard satisfies the selected $\\pm0.05$-PPL equivalence criterion on every checkpoint."],"supporting_citations":[{"why":"Defines the QuaRot W4A4KV4 rotated-quantisation baseline whose online full-head Hadamard $R_3$ is the comparison configuration for every experiment.","marker":"(Ashkboos et al., 2024b)"},{"why":"Supplies the per-pair RoPE-commuting rotation family that the paper's Lemma 2 completes with the converse characterisation and then evaluates.","marker":"(van Breugel et al., 2025)"},{"why":"Provides the whip calibration objective and the R1–R4 rotation procedure whose released defaults the experiments reproduce exactly.","marker":"(Shao et al., 2025)"},{"why":"Defines RoPE and the frequency-pair decomposition of each head that motivates the pairwise rotation's mixing support.","marker":"(Su et al., 2024)"},{"why":"Supplies the GPTQ weight quantisation used for the W4A4KV4 weight side of the evaluated set-up.","marker":"(Frantar et al., 2023)"},{"why":"Supplies the interval-based TOST procedure behind the selected ±0.05-PPL equivalence criterion for pairwise+Hadamard.","marker":"(Lakens, 2017)"}],"fun_headline_variants":["Exact local optimum still loses to full-head mixing","Optimal pairwise rotation can't beat full-head mixing","RoPE-aligned rotation's exact optimum falls short","Why the variance-optimal pairwise rotation fails","Local variance optimality doesn't fix 4-bit quantisation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the head-shared parameterisation: one rotation angle per layer and frequency pair, broadcast to every attention head, which is the only subfamily the experiments test, while the paper's own centraliser characterisation also permits head-dependent angles that remain unmeasured.","fun_headline_variants_meta":{"raw":{"variants":["Exact local optimum still loses to full-head mixing","Optimal pairwise rotation can't beat full-head mixing","RoPE-aligned rotation's exact optimum falls short","Why the variance-optimal pairwise rotation fails","Local variance optimality doesn't fix 4-bit quantisation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1553,"prompt_tokens":1326,"completion_tokens":227,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":942,"completion_tokens_details":{"reasoning_tokens":151}},"tokens_in":942,"tokens_out":227,"duration_ms":2938,"temperature":1.0,"reasoning_tokens":151,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:46:52.870424+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the exact W4A4KV4 protocol with head-dependent per-pair angles $\\phi_{h,k}$, which the paper's Lemma 2 allows: if any such configuration matches or beats the full-head Hadamard baseline on WikiText-2 perplexity, the negative result is an artefact of the head-shared constraint rather than of RoPE-aligned pairwise rotation itself. Alternatively, optimise the rotation against the quantiser's actual tokenwise group range statistic instead of the position-averaged second moment; if perplexity then reaches the Hadamard level, the surrogate mismatch — not the two-channel support — is the operative cause.","supporting_citations":[],"review_version":1}