{"id":"3bba9a22-baa5-4382-bd44-fa1f9b5a57e5","arxiv_id":"2505.23651","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"HDRQ adds loss-surface flattening and source-weight distance regularization to post-training quantization, improving how well quantized domain-adapted models merge.","lead":"This paper presents a quantization method that keeps separately adapted neural networks easy to merge. It matters because deploying many specialized models on small devices requires quantizing them without destroying the benefits of weight averaging.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing the QDrop+same-regularizers control makes the attribution of merging gains to Hessian/distance regularization unproven, so the paper's central claim of a merge-friendly quantization redesign is conditional at best.","rationale":"The paper's empirical setting is plausible and the headline improvement is nontrivial at low precisions, which is genuine evidence for the practical recipe. However, the verification gap is that the controlled experiments needed to support the paper's specific attribution are absent. The paper's ablations (Table 3) only stack components on top of QDrop, never adding a single component to QDrop in isolation or testing the components in a 2x2 design. The domain-adaptation analysis in §3.3 is also heuristic: the constant k in Eq. (7) is unspecified, the 'gradient is a scaled version of the interpolation direction' claim is asserted from the one-basin assumption rather than checked, and Eq. (6)'s M-Lipschitz condition is never measured. The reader identified these theory-side weaknesses (unmeasured M, unverified basin/gradient-scaling assumptions, Eq. (4) max-split) as the weakest assumption; my stress-test agrees with that assessment but identifies the ablation/comparison gap as the more empirically decisive issue, since the empirical method could still work even if the theory is imperfect. I would keep the verdict CONDITIONAL: the paper should release code/checkpoints, add the QDrop+regularizer control, report variances (Table 1 reports means over 30 samples but no deviations), and either tighten Eq. (7) or explicitly label the analysis as heuristic. The central claim could hold, but it is not yet proven that the named components, rather than the engineering recipe (block-wise reconstruction details, noise-sampling rounding, training schedule), drive the gains.","tokens_in":14581,"tokens_out":1990,"duration_ms":16423,"concrete_test":"Run the semantic-segmentation W4A4 and W3A3 merges with four configurations: (1) exact QDrop, (2) QDrop + the same weight-distance regularization term (Eq. 10 with lambda=5e-2), (3) QDrop + the same noise-sampling rounding but no distance term, (4) full HDRQ. If (4) does not beat (2) and (3) by more than the reported 4.21 mIoU gap over QDrop in the W4A4 G->C/I merge, the paper's attribution of merging gains to the Hessian and distance regularizers is not supported, and the theoretical motivation does not explain the headline result.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central practical claim (HDRQ improves merged-model accuracy at low precision, e.g., 4.21 mIoU over QDrop at W4A4) rests on an attribution claim: that the gains come from (i) noise-based Hessian regularization and (ii) weight-distance regularization from the source, both inspired by the error-barrier analysis in Sections 3.1-3.3. The ablations in Table 3 are incremental from QDrop, but the components are not independently varied on each side. Critical missing baselines: (1) QDrop + distance regularization only; (2) QDrop + noise-based quantization only (i.e., without Hessian/distance logic); (3) QDrop + both modifications, to confirm the effects are additive and that the proposed 'advanced' noise-sampling rounding, held out of Table 3, is not the real driver. Without these, the ablation shows only that HDRQ's recipe beats QDrop, not that the Hessian and distance regularizers—the theory's load-bearing elements—are what cause the merging benefit. Additionally, the weight-distance regularizer (Eq. 10) is applied per layer to the source weight, which is a surrogate for the inter-target distance; the bound is valid but the regularization strength lambda=5e-2 is fixed and unsearched across tasks, and the Office-Home results show HDRQ trailing QDrop in several W8A8/W8A4 cells (e.g., R->A,C,P W8A8: 67.46 vs 67.91; C->R,A,P W8A8: 74.58 vs 74.79), so the 'comparable or better single-model performance' claim is not uniformly supported. The most load-bearing unverified step is the attribution; it can be settled by a controlled head-to-head that adds the same two regularizers to QDrop and measures the merged-model harmonic mean at W4A4/W3A3.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies post-training quantization of separately domain-adapted models that are subsequently merged for multi-target domain adaptation. It proposes HDRQ, which combines noise-based Hessian regularization, a weight-distance regularization pulling each quantized model toward the source weights, and a noise-sampling rounding procedure designed to resolve rounding ambiguity during merging. The authors present a theoretical error-barrier analysis (Section 3) motivating these components and evaluate on GTA-to-Cityscapes/IDD semantic segmentation and Office-Home classification, reporting that HDRQ improves merged-model accuracy at low precision (e.g., +4.21 mIoU over QDrop at W4A4) while keeping single-model accuracy roughly comparable.","tokens_in":14913,"tokens_out":6876,"duration_ms":67938,"significance":"If substantiated, the paper identifies a practically important design axis for post-training quantization: current PTQ methods optimize per-model accuracy but ignore the compatibility of quantized models with subsequent weight merging. The empirical package is broad, using standard benchmarks and baselines, and the incremental ablation in Table 3 gives initial evidence that each proposed component contributes. However, the theoretical derivation as written is not rigorous, and the attribution of the headline merging gains to the two regularizers is not fully isolated from the advanced noise-sampling rounding in the main segmentation experiment. The work is a promising step rather than a settled result.","major_comments":[{"comment":"The displayed equality is not valid because the maximum of a sum is not generally equal to the sum of the maxima; only an inequality max[f+g] <= max[f] + max[g] holds without additional conditions. The subsequent minimization argument in Eq. (5) depends on this decomposition, so the theoretical analysis is not sound as written. Please rewrite the step as an inequality or prove the required exchange-of-maxima condition, and state how this affects the conclusions drawn from Eqs. (5)-(7).","section":"Section 3.1, Eq. (4)"},{"comment":"The scalar k is introduced as being proportional to the distance between θλ and θ2, but it is never defined, estimated, or bounded. The assertion that the Jacobian term ∇_θ L1(θ2) is a scaled copy of the interpolation direction is an unverified modeling assumption. Likewise, Eq. (6) invokes an M-Lipschitz Hessian without any attempt to estimate M or check the assumption on the actual adapted models. These assumptions are load-bearing because they justify both the distance regularization and the Hessian regularization. I would accept a clearly labeled heuristic motivation, but the current text presents these as derivation; please either measure/validate the assumptions (for example, gradient alignment or Hessian Lipschitz constants along the interpolation path) or explicitly downgrade the theoretical claims.","section":"Section 3.3, Eq. (7)"},{"comment":"The attribution of the headline merging gain to the Hessian and distance regularizers is not fully established because the advanced noise-sampling rounding of Section 4.3 is not controlled for on the semantic segmentation task. The W4A4 +4.21 mIoU result is reported after merging 30 sampled weights, and the cosine-similarity-based selection is HDRQ-specific, but the text does not state whether the QDrop and BRECQ baselines also receive the same noise-sampling procedure. If the sampling is applied only to HDRQ, the comparison is unfair; if it is applied to all methods, that should be stated explicitly. In addition, Table 3 (Office-Home) is conducted in a setting where the paper says sampling is not applied, so it does not control for this component in the task where the headline gain is obtained.","section":"Sections 5.1-5.3, Tables 1-3"},{"comment":"The summary claims that HDRQ 'consistently outperforms' previous approaches and maintains 'comparable or superior' single-model performance, but the tables do not uniformly support this. At W8A8 in Table 2, the merged harmonic mean for HDRQ is below QDrop for all four source configurations (67.46 vs 67.91, 68.10 vs 68.75, 74.58 vs 74.79, 64.51 vs 64.86). In Table 1, the W4A4 single-model mIoU is lower for HDRQ for both targets (58.23 vs 58.92 on G->C and 48.68 vs 49.44 on G->I). The paper should report variance (standard deviations or multiple seeds) and either temper these broad claims or discuss the trade-off between single-model accuracy and merging quality.","section":"Section 5.2, Table 2; Section 5.1, Table 1"},{"comment":"The experimental comparison is reported as point estimates without error bars, seed information, or standard deviations. This is especially problematic for the small gaps in Table 1 (for example, 61.54 vs 61.32 at W6A6) and for the semantic segmentation results that average over 30 sampled weights. Without variance estimates, the reader cannot assess whether the reported differences are statistically meaningful. Please add multiple independent runs or at least report the standard deviation of the merging metric across the sampled weights, and specify the number of independent quantization runs per condition.","section":"All experimental tables"}],"minor_comments":[{"comment":"The W8A4 entry appears as '67.464.85', which looks like two numbers run together; please correct the formatting and re-verify the corresponding values.","section":"Table 2, QDrop row for R->A,C,P"},{"comment":"The text says the results are presented in 'Section 4.1', but the semantic segmentation results appear in Table 1 of Section 5.1; the cross-reference is wrong.","section":"Section 5.1"},{"comment":"The merging formula for integer representations should explicitly state the assumed merged step size, such as (Δ1+Δ2)/2; otherwise the expression I_merged = floor((I1*Δ1 + I2*Δ2)/(Δ1+Δ2)) is not self-explanatory.","section":"Eq. (11)"},{"comment":"The axis labels v1 and v2 are not defined in the caption or text; please specify what these directions are (for example, principal components of the weight space).","section":"Figure 2"},{"comment":"The sentence 'we minimize the second-order term in (7)' refers to Eq. (7), but Eq. (7) is introduced later in Section 3.3; please correct the cross-reference or renumber the equations.","section":"Section 3.2"},{"comment":"No code or repository is mentioned; releasing the implementation would substantially improve reproducibility, especially given the many implementation details in the quantization pipeline.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper extends the authors' prior NIPQ noise-based quantization work; the genuinely new part is the merging-compatibility framing and the distance regularization to the source weights. The theoretical section currently overreaches, and the missing control for the advanced noise-sampling rounding is the main empirical gap. With additional controls and more careful claims this could become a solid contribution to the PTQ and model-merging literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the empirical headline is real: at W4A4 semantic segmentation, merging models quantized with HDRQ beats QDrop-merged models by 4.21 mIoU, and the gains are consistent at low bit-widths across two tasks. Second, the theory that motivates the method does not hold up line by line, and the ablation is one baseline short of supporting the central attribution claim. If you read it expecting a rigorous proof that Hessian-plus-distance regularization is the right way to make PTQ merge-friendly, you'll be disappointed; if you read it as an engineering paper proposing a practical recipe, it's mostly convincing.\n\nWhat's genuinely new: framing PTQ as a merging-compatibility problem in MTDA, and adding a source-distance regularizer plus a noise-sampling rounding trick to a noise-based PTQ baseline. The paper is honest that noise-based quantization itself is prior work (Baskin, Défossez, Shin, Lin), and it cites its own NIPQ for the Hessian-regularization mechanism. The experiments are extensive for a PTQ paper: two tasks, multiple bit settings, per-class segmentation results, and an ablation. The W4A4 segmentation result is the kind of concrete, practical win that justifies publication if it survives scrutiny.\n\nThe soft spots are real but not fatal. The derivations in Section 3 are loose: Eq. (4) splits a max over a sum into a sum of maxima without justification, Eq. (7) introduces an unspecified scalar k, and the basin/gradient-scaling assumptions are asserted, not checked. The appendix doesn't report error bars or release code, which matters because the segmentation numbers come from 30 sampled weights. The ablation in Table 3 is incremental from QDrop with both components added at once; it never runs QDrop plus distance regularization only, or QDrop plus noise-based quantization only, so you can't tell whether the 4.21-point gain is due to the two advertised regularizers or to the advanced noise-sampling rounding held out of the table. Also, the 'comparable or superior' claim is not uniform: in several Office-Home W8A8/W8A4 cells HDRQ trails QDrop by 0.2-0.5 points (e.g., R->A,C,P W8A8: 67.46 vs 67.91). None of this undermines the main practical claim—HDRQ is a better recipe at low precision—but it does mean the paper's theoretical framing and attribution are conditional at best.\n\nWho this is for: researchers working on PTQ or on efficient deployment of domain-adapted models. It's a serious engineering contribution that deserves a proper referee; the right outcome would be major revision asking for the missing QDrop+regularizer baselines, error bars, code, and a cleaned-up derivation. I'd bring it to reading group, but I wouldn't cite it as evidence for the theory, only for the empirical recipe.","headline":"HDRQ is a plausible, well-motivated engineering contribution with a useful empirical result, but its theoretical derivation is shaky and the ablation doesn't pin the gains to the two advertised regularizers.","tokens_in":15486,"tokens_out":942,"would_cite":false,"duration_ms":10804,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that quantization noise degrades model merging, and that HDRQ — a quantization scheme that flattens loss surfaces and anchors weights to the source — restores merging quality, gaining 4.21 mIoU at W4A4.","keywords":["post-training quantization","model merging","multi-target domain adaptation","error barrier","Hessian regularization","noise-based quantization","weight averaging","loss landscape flattening"],"falsifier":"Estimate the Hessian Lipschitz constant $M$ along the interpolation path between two domain-adapted models and check whether the gradient of one domain's loss at the other model is proportional to the interpolation vector; if the proportionality fails, the theoretical motivation for the distance regularizer collapses. A simpler check: merge two models after randomly permuting one model's weight indices to break the shared-basin assumption — if HDRQ still improves the merged result, its benefit is empirical rather than basin-based.","tokens_in":14364,"feed_emoji":"🧩","tokens_out":13403,"duration_ms":98524,"temperature":0.7,"pith_summary":"The paper identifies quantization noise as a hidden cause of poor model merging. When two domain-adapted models are each quantized before being averaged into one multi-target model, the discrete weight perturbations inflate the error barrier between them and the merged model loses accuracy. The authors propose HDRQ, a post-training quantization method that flattens the loss surface around each model by injecting quantization noise, pulls each quantized model back toward the shared source weights with an $\\ell^2$ distance regularizer, and resolves rounding ambiguity during merging with noise sampling. In W4A4 semantic segmentation, merging models quantized with HDRQ outperforms merging models quantized with QDrop by 4.21 mIoU, while single-model accuracy stays comparable or better.","feed_headline":"Quantization noise breaks model merging; HDRQ fixes it","feed_subtitle":"Flattens loss surfaces, anchors weights to the source, and lifts merged-segmentation mIoU by 4.21.","key_machinery":"The error barrier — the largest loss increase along the linear interpolation between two adapted models, relative to their mean loss — is the central object. HDRQ's theoretical story is that quantization adds uniform noise to weights, and a second-order Taylor expansion turns the barrier into terms controlled by the Hessians at the two endpoints and the merged point. Regularizing the Hessian (through injected noise), reducing the distance between the two models (through an $\\ell^2$ pull toward the source), and resolving the rounding-parity ambiguity (through cosine-filtered noise sampling) each target one term in that expansion, with the Hessian's M-Lipschitz continuity providing the link between endpoint curvature and merged-point curvature.","core_discovery":"The central claim is that quantization should be designed with a future merge in mind, not only with single-model accuracy in mind. Working in the setting of training-free multi-target domain adaptation, where separately fine-tuned models are fused by weight averaging, the paper shows that quantization noise raises the error barrier between the two adapted weight points, and that this rise can be countered at quantization time. HDRQ does this with three linked mechanisms: noise-based quantization whose expected loss includes a Hessian-penalizing second-order term, thereby flattening the loss landscape; a distance regularizer that keeps each quantized model close to the shared source weight, which by the triangle inequality keeps the two target models close to each other; and a noise-sampling rounding step, filtered by cosine similarity, that removes the ambiguity of rounding an odd midpoint when averaging two quantized weight grids. The experiments across semantic segmentation and Office-Home show the merged model's harmonic mean accuracy improves most at low bit-widths, which is exactly where quantization noise is largest.","pith_inferences":["Beyond the paper's setting, the same recipe — flatten the Hessian and anchor to a shared starting point — could plausibly improve merging for other post-hoc weight operations such as task arithmetic or model soups, though the paper only demonstrates it for domain-adapted models under simple weight averaging.","A testable extension the paper does not pursue is to measure the Lipschitz constant $M$ and the gradient-scaling scalar $k$ directly; small values would make the theory's bounds tight, while large values would leave the method's empirical success unexplained by the stated analysis.","The cosine-similarity-filtered sampling suggests a general principle for stochastic merges: when the merged weights are ambiguous, choose the sample whose merged direction best preserves the original interpolation geometry. This could be tried as a drop-in fix for any midpoint weight averaging, with or without quantization."],"forward_implications":["Post-training quantization should be benchmarked on merging outcomes in addition to single-model accuracy, since the two metrics diverge sharply at low bit-widths.","The gap between HDRQ and QDrop grows from 1.19 to 4.21 mIoU as precision drops from W6A6 to W4A4, so merge-friendliness matters most where quantization noise is largest.","Noise-based quantization, used in prior single-model post-training quantization work, gains a new role: it can deliberately flatten the loss surface to make later weight averaging across models succeed.","The triangle-inequality argument means keeping quantized models close to their shared source is a sufficient proxy for keeping them close to each other, without ever seeing the other target domain."],"supporting_citations":[{"why":"Supplies the training-free weight-averaging merging method for multi-target domain adaptation and the shared-basin assumption HDRQ builds on.","marker":"Li et al., 2024"},{"why":"Defines the error barrier that the paper extends to quantized weights.","marker":"Frankle et al., 2020"},{"why":"Establishes the view of quantization noise as additive uniform perturbation, the basis for noise-based Hessian regularization.","marker":"Baskin et al., 2021"},{"why":"Introduces pseudo quantization noise, the mechanism behind the noise-sampling rounding and Hessian regularization.","marker":"Défossez et al., 2022"},{"why":"Provides the noise-proxy-based quantization framework that HDRQ draws on for stable noise-based weight updates.","marker":"Shin et al., 2023"},{"why":"QDrop serves as the primary post-training quantization baseline and supplies the block-wise reconstruction with partial dropout that HDRQ extends.","marker":"Wei et al., 2022"},{"why":"BRECQ supplies the block-wise reconstruction backbone and is one of the main baselines compared for merging quality.","marker":"Li et al., 2021"},{"why":"Bit-shrinking links noise-based quantization with sharpness-aware scheduling, the immediate antecedent for using noise to flatten the loss surface during post-training quantization.","marker":"Lin et al., 2023"}],"fun_headline_variants":["Quantization raises merge barriers; HDRQ flattens them","HDRQ: quantize for merging, not just for accuracy","Low-bit quantization wrecks merges; HDRQ smooths the path","Merge-friendly quantization: HDRQ anchors and flattens loss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the loss landscape between the two adapted models is smooth enough that the Hessian changes slowly (an M-Lipschitz bound) and that the models truly occupy one shared basin, so one domain's gradient at the other model is a scaled copy of the interpolation direction; the paper does not measure either quantity, and if either fails the distance regularizer's theoretical justification loses its footing.","fun_headline_variants_meta":{"raw":{"variants":["Quantization raises merge barriers; HDRQ flattens them","HDRQ: quantize for merging, not just for accuracy","Low-bit quantization wrecks merges; HDRQ smooths the path","Merge-friendly quantization: HDRQ anchors and flattens loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000185,"raw_usage":{"total_tokens":1301,"prompt_tokens":906,"completion_tokens":395,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":318}},"tokens_in":522,"tokens_out":395,"duration_ms":3923,"temperature":1.0,"reasoning_tokens":318,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:40:50.757097+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Estimate the Hessian Lipschitz constant $M$ along the interpolation path between two domain-adapted models and check whether the gradient of one domain's loss at the other model is proportional to the interpolation vector; if the proportionality fails, the theoretical motivation for the distance regularizer collapses. A simpler check: merge two models after randomly permuting one model's weight indices to break the shared-basin assumption — if HDRQ still improves the merged result, its benefit is empirical rather than basin-based.","supporting_citations":[{"cited_title":"K., Roy, D., and Carbin, M","cited_arxiv_id":null,"evidence_quote":"Defines the error barrier that the paper extends to quantized weights."},{"cited_title":"M., and Mendelson, A","cited_arxiv_id":null,"evidence_quote":"Establishes the view of quantization noise as additive uniform perturbation, the basis for noise-based Hessian regularization."},{"cited_title":"Nipq: Noise proxy-based integrated pseudo-quantization","cited_arxiv_id":null,"evidence_quote":"Provides the noise-proxy-based quantization framework that HDRQ draws on for stable noise-based weight updates."},{"cited_title":"QD rop: Randomly dropping quantization for extremely low-bit post-training quantization","cited_arxiv_id":null,"evidence_quote":"QDrop serves as the primary post-training quantization baseline and supplies the block-wise reconstruction with partial dropout that HDRQ extends."},{"cited_title":"Bit-shrinking: Limiting instantaneous sharpness for improving post-training quantization","cited_arxiv_id":null,"evidence_quote":"Bit-shrinking links noise-based quantization with sharpness-aware scheduling, the immediate antecedent for using noise to flatten the loss surface during post-training quantization."}],"review_version":1}