{"id":"0958bf4a-0a31-4cdf-95eb-0c388bae6d8b","arxiv_id":"2607.07494","paper_version":1,"verdict":"CONDITIONAL","confidence":"UNKNOWN","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":4,"one_line_summary":"Transforming gradients into K-FAC-based coordinates before FP8 quantization reduces communication error and improves downstream task preservation over Euclidean FP8, with a 7.6% end-to-end speedup on 64 GH200 GPUs.","lead":"GIFT transforms gradients into a geometry-aware coordinate system before FP8 quantization to reduce communication distortion in LLM pretraining. A smart generalist might read it because it offers a principled way to cut distributed training costs without degrading model quality.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The decisive evidence—downstream task comparison—rests on single-seed runs with a coarse 'wins vs FP32' metric that masks per-task regressions and lacks significance testing.","rationale":"The reader's verdict of CONDITIONAL already captures the right posture: the geometric principle is sound and the round-trip analysis is convincing, but the downstream evidence is not yet conclusive. My concern is more load-bearing than the reader's chosen weakest_assumption (K-FAC factor staleness). The K-FAC staleness question is valid but secondary: the round-trip error analysis already shows the factors work well at the tested refresh interval, and the geometric argument is clean. The real soft spot is that the paper's own 'decisive test' (downstream tasks) lacks the statistical power to support the claim it is asked to bear. A single seed with a binary win-count metric over 14 tasks, where several tasks show GIFT worse than Euclidean, does not establish that geometry-aware communication reliably improves downstream quality. The verdict remains CONDITIONAL because the mechanism is promising and the systems design is reasonable, but the key empirical claim needs replication. I set agreement_with_reader to 'partial' because the reader identified this weakness in their rationale but chose K-FAC staleness as the weakest_assumption instead.","tokens_in":17473,"tokens_out":1989,"duration_ms":100020,"concrete_test":"Re-run the 600M Llama pretraining with GIFT and Euclidean FP8 using at least 3 random seeds each. For each of the 14 downstream tasks, compute mean ± std and a paired comparison. If the 'wins vs FP32' count for GIFT drops below 6/14 on average, or if the per-task differences between GIFT and Euclidean are not statistically significant (p > 0.05) on a majority of tasks, the downstream advantage claim weakens substantially. This directly tests whether the 7-vs-5 margin is signal or noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's own logic chain is: (1) round-trip FP8 error analysis shows geometry-aware coordinates reduce reconstruction error (Tables I–II, mechanistically sound), and (2) this translates to better downstream task preservation (Table III, the 'decisive test' per §VI.D). Part (2) is the load-bearing link for practical relevance, and it is statistically fragile. The headline metric 'wins vs FP32' is a binary count over 14 tasks: GIFT wins 7/14 vs Euclidean's 5/14 at 600M. But this masks that GIFT substantially underperforms Euclidean on several individual tasks (e.g., 600M CB-ACC: GIFT 0.3036 vs Euclidean 0.4107; RCD-F1: GIFT 0.8197 vs Euclidean 0.8271; WINO: GIFT 0.5257 vs Euclidean 0.5320). With no error bars, no multiple seeds, and only 14 binary comparisons, the 7-vs-5 margin could easily be noise. The paper acknowledges this in §VI.G ('additional seeds and a larger benchmark suite' needed), but the central claim of downstream improvement is not yet established at conventional statistical confidence. The round-trip error reduction is real and well-demonstrated; whether it produces a reliable downstream advantage is the untested step.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This paper proposes GIFT, a geometry-informed coordinate transformation for FP8 gradient communication in distributed LLM pretraining. The core idea is to whiten gradients using an input-side K-FAC factor before FP8 quantization and communication, so that quantization error is distributed more isotropically across gradient directions. The authors progressively simplify the full two-sided K-FAC transform to: (1) input-side only, (2) rank-32 low-rank approximation, and (3) selective deployment on the 13 most numerically vulnerable fc2 layers. They evaluate on Llama-300M and Llama-600M with the Muon optimizer, showing up to 67.4% reduction in FP8 round-trip reconstruction error over the Euclidean baseline, a 7.6% end-to-end pretraining speedup over FP32 on 64 GH200 Superchips, and a more favorable downstream task preservation profile (7/14 tasks vs FP32) compared to direct Euclidean FP8 (5/14 tasks).","tokens_in":17733,"tokens_out":1809,"duration_ms":277240,"significance":"The paper addresses a practically important problem: gradient communication is a well-known scaling bottleneck in LLM pretraining, and naive FP8 quantization introduces direction-dependent distortion under anisotropic gradient geometry. The coordinate-transformation perspective is novel relative to prior compression work (QSGD, PowerSGD, 1-bit Adam, SDP4Bit) that modifies quantization rules, sparsity, or communication patterns. The round-trip error analysis (Tables I–II) is mechanistically clean and provides a falsifiable, parameter-light justification for the input-side simplification. The systems evaluation on real GH200 hardware with end-to-end pretraining and a concrete 7.6% speedup measurement is a strength. The method is complementary to existing optimizers and communication collectives, which is a practically useful design property. However, the downstream task evaluation that supports the central practical claim is statistically fragile (single seed, 14 binary comparisons, no significance testing), which limits the strength of the evidence for the headline claim.","major_comments":[{"comment":"§VI.D, Table III: The central practical claim—that geometry-aware communication improves the downstream preservation profile over Euclidean FP8—rests on single-seed runs evaluated via a binary 'wins vs FP32' count over 14 tasks. The 7-vs-5 margin (600M) and 7-vs-4 margin (300M) could easily arise from noise given the per-task variance visible in the table (e.g., 600M CB-ACC: GIFT 0.3036 vs Euclidean 0.4107; RCD-F1: GIFT 0.8197 vs Euclidean 0.8271; WINO: GIFT 0.5257 vs Euclidean 0.5320). Without error bars, multiple seeds, or any significance test, the downstream advantage is not established at conventional confidence. The paper acknowledges this in §VI.G ('additional seeds and a larger benchmark suite' needed), but the claim in the abstract and conclusion is stated without this qualification. At minimum, the paper should either (a) run at least 3 seeds and report mean ± std for each task","section":null},{"comment":"§V.E, Algorithm 1: The K-FAC input-side factor A is updated every K=50 steps, but no sensitivity analysis is provided for this choice. The reader's concern about staleness is legitimate: if the curvature changes faster than the 50-step refresh interval, the whitening transform may become stale and the fidelity benefit could degrade. The paper should either provide an ablation over K (e.g., K ∈ {10, 25, 50, 100}) or cite evidence that the input-side activation statistics are stable over 50-step windows for the models and training recipes studied.","section":null},{"comment":"§V.C, Figure 3: The layer-selection procedure (top 13 vulnerable fc2 layers) is profiled on the 600M model during the first 100 training steps. It is unclear whether this vulnerability ranking is stable throughout training or whether it reflects only early-training dynamics. If the vulnerable layer set shifts as training progresses, the fixed selection may miss layers that become vulnerable later. The paper should discuss whether the profiling window is representative of the full training run, or ideally re-profile at multiple training checkpoints to verify stability.","section":null},{"comment":"§VI.D, Table III (300M row): The full K-FAC variant wins 6/14 tasks vs FP32, while the selective GIFT design wins 7/14. The paper frames this as supporting the selective design, but the 6-vs-7 difference over 14 tasks is within noise. More importantly, on several individual tasks full K-FAC substantially outperforms GIFT (e.g., CB-ACC: full K-FAC 0.4107 vs GIFT 0.2321; MUL-RC: full K-FAC 0.5140 vs GIFT 0.4420). The claim that selective deployment is superior to full K-FAC is not well-supported by this data and should be softened or supported by additional evidence.","section":null}],"minor_comments":[{"comment":"§I: The abstract states 'improving the downstream task preservation profile over direct Euclidean FP8 communication.' Given the statistical fragility discussed above, this should be qualified (e.g., 'in our single-seed experiments, improving...').","section":null},{"comment":"§II.A: The notation 'g^(r)' in Eq. (1) uses a superscript that could be confused with exponentiation; consider using a subscript or explicit indexing.","section":null},{"comment":"§IV.A, Eq. (14): The notation L^{-⊤}_A is used but not explicitly defined as (L_A^{-1})^⊤. A brief clarification would help readers unfamiliar with this convention.","section":null},{"comment":"§V.B, Table II: The 'block-A-32' and 'block-A-64' methods are mentioned but not clearly defined in the text. A brief description of the block-diagonal approximation structure would improve reproducibility.","section":null},{"comment":"§VI.A: The choice of different sequence lengths for the two models (4096 for 300M, 2048 for 600M) is explained as a memory constraint, but this means the two models are not directly comparable. This should be noted more prominently when interpreting cross-model trends.","section":null},{"comment":"§VI.E: The memory overhead for 600M (8.98%) is notably higher than for 300M (3.33%). The paper should briefly explain why the overhead scales non-linearly with model size (e.g., is it because the number of selected layers or the factor dimensions scale differently?).","section":null},{"comment":"Table III: The 'LAMB-STD' column header appears to differ in format from other headers (period vs hyphen). Consistency would improve readability.","section":null},{"comment":"§VI.D: The paper states 'BF16 matches FP32 on 7 out of 14 tasks, but its absolute task values are slightly lower overall.' This claim of 'slightly lower overall' should be supported by an aggregate metric (e.g., average across tasks) rather than stated qualitatively.","section":null},{"comment":"References: The Muon optimizer [8] is cited as a web page rather than a peer-reviewed publication. If a preprint or formal report is available, it should be cited accordingly; if not, the citation format should note it as a software/technical report.","section":null}],"recommendation":"major_revision","confidential_remarks":"The round-trip error analysis is genuinely solid and the coordinate-transformation idea is a clean conceptual contribution. The main concern is that the downstream evaluation—the load-bearing evidence for practical relevance—is too statistically fragile to support the headline claim as stated. With 2-3 seeds and appropriate significance testing (or at minimum, honest qualification of the claim), this paper could be a solid contribution. The authors' own acknowledgment of the limitation in §VI.G suggests they are aware of this. I would encourage the editor to give them the opportunity to strengthen the evaluation rather than reject outright, as the mechanistic analysis and systems implementation are well-executed. The paper fits the journal's scope in distributed systems and LLM training infrastructure."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive report. The referee raises four major points, all of which concern the strength of empirical evidence supporting our claims: (1) the downstream task evaluation lacks multiple seeds and significance testing; (2) no sensitivity analysis for the K-FAC refresh interval K=50; (3) the layer-vulnerability profiling window may not be representative of the full training run; and (4) the claim that selective GIFT is superior to full K-FAC is not well-supported by the 6-vs-7 task count. We agree with the substance of all four points and will revise the manuscript accordingly. Specifically, we will run additional seeds with mean±std reporting, add a K-sensitivity ablation, re-profile vulnerability at multiple checkpoints, and soften the selective-vs-full-K-FAC claim. We also note that the mechanistic round-trip error analysis (Tables I–II) and the systems-level speedup measurement are independent of these concerns and remain valid as stated.","responses":[{"response":"The referee is correct. The downstream evaluation as presented is single-seed with binary win counting and no significance testing, which is insufficient to establish the claimed advantage at conventional confidence. We will address this by running at least 3 seeds for each method (FP32, Euclidean FP8, GIFT) on both model sizes and reporting mean ± std for each task. We will also apply paired significance tests (e.g., Wilcoxon signed-rank) across tasks. The abstract and conclusion will be revised to include appropriate qualifications (e.g., 'in our experiments' or 'under the seeds tested') rather than stating the downstream advantage as an unqualified fact. We note that the mechanistic round-trip error analysis (Tables I–II) showing 67.4% reduction in reconstruction error is independent of the downstream evaluation and is not affected by this concern. The 7.6% end-to-end speedup is a systems measurement that does not depend on seed variance.","revision_made":"yes","referee_comment":"§VI.D, Table III: The central practical claim—that geometry-aware communication improves the downstream preservation profile over Euclidean FP8—rests on single-seed runs evaluated via a binary 'wins vs FP32' count over 14 tasks. The 7-vs-5 margin (600M) and 7-vs-4 margin (300M) could easily arise from noise given the per-task variance visible in the table. Without error bars, multiple seeds, or any significance test, the downstream advantage is not established at conventional confidence. The paper acknowledges this in §VI.G but the claim in the abstract and conclusion is stated without this qualification."},{"response":"We agree that a sensitivity analysis for K is needed and absent from the current manuscript. We will add an ablation over K ∈ {10, 25, 50, 100} measuring both the round-trip reconstruction error (as in Tables I–II) and, if compute permits, validation loss trajectories. Our intuition is that input-side activation second-order statistics (A = E[aa^T]) are relatively stable over 50-step windows because they depend on input distributions rather than rapidly changing gradient dynamics, but this should be verified empirically rather than asserted. If the ablation shows that shorter intervals improve fidelity meaningfully, we will report the cost-quality tradeoff explicitly. We will also add a brief discussion of why input-side statistics may be more stable than output-gradient-side statistics, which was part of our motivation for the input-only simplification.","revision_made":"yes","referee_comment":"§V.E, Algorithm 1: The K-FAC input-side factor A is updated every K=50 steps, but no sensitivity analysis is provided for this choice. The reader's concern about staleness is legitimate: if the curvature changes faster than the 50-step refresh interval, the whitening transform may become stale and the fidelity benefit could degrade. The paper should either provide an ablation over K or cite evidence that the input-side activation statistics are stable over 50-step windows."},{"response":"This is a valid concern. The current profiling is limited to the first 100 training steps, and we have not verified that the vulnerability ranking is stable throughout training. We will re-profile the vulnerability ranking at multiple checkpoints (e.g., early, mid, and late training) for the 600M model and report whether the top-13 layer set changes. If the set is stable, we will state this explicitly with the supporting evidence. If the set shifts, we will discuss the implications: (a) a fixed selection may be suboptimal for later training, and (b) periodic re-selection could be a practical extension. We will also add an explicit caveat in §V.C noting that the profiling window is early-training and that the stability assumption is empirically tested rather than theoretically guaranteed.","revision_made":"yes","referee_comment":"§V.C, Figure 3: The layer-selection procedure (top 13 vulnerable fc2 layers) is profiled on the 600M model during the first 100 training steps. It is unclear whether this vulnerability ranking is stable throughout training or whether it reflects only early-training dynamics. If the vulnerable layer set shifts as training progresses, the fixed selection may miss layers that become vulnerable later. The paper should discuss whether the profiling window is representative of the full training run, or ideally re-profile at multiple training checkpoints to verify stability."},{"response":"The referee is correct on both counts. The 6-vs-7 difference is within noise, and on several individual tasks full K-FAC outperforms selective GIFT. The current framing overstates the evidence. We will soften the claim from 'supporting the selective design' to something like 'the selective design achieves a comparable downstream profile to full K-FAC while incurring lower computational cost, which is the primary practical motivation for the selective approach.' The argument for selective deployment is better justified on cost grounds (lower overhead) than on downstream-quality grounds, and we will make this distinction explicit. We will also add the multi-seed evaluation for the full K-FAC variant on 300M, which will provide a fairer comparison. If the multi-seed results show no significant difference between full K-FAC and selective GIFT on downstream tasks, we will frame the selective design purely as a cost-reduction measure that preserves the fidelity benefit, not as a quality improvement.","revision_made":"yes","referee_comment":"§VI.D, Table III (300M row): The full K-FAC variant wins 6/14 tasks vs FP32, while the selective GIFT design wins 7/14. The paper frames this as supporting the selective design, but the 6-vs-7 difference over 14 tasks is within noise. More importantly, on several individual tasks full K-FAC substantially outperforms GIFT (e.g., CB-ACC: full K-FAC 0.4107 vs GIFT 0.2321; MUL-RC: full K-FAC 0.5140 vs GIFT 0.4420). The claim that selective deployment is superior to full K-FAC is not well-supported by this data and should be softened or supported by additional evidence."}],"tokens_in":17803,"tokens_out":1490,"duration_ms":123531,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"The core idea is genuinely new: use K-FAC factors not as an optimizer preconditioner but as a coordinate transform for FP8 gradient communication. The derivation from the Fisher metric ball to a Euclidean ball (Eqs. 7–13) is clean and correct. The round-trip error analysis (Tables I–II) is the strongest part — the input-side simplification is well-justified, and the progressive ablation from full two-sided K-FAC down to input-only rank-32 is done with the right controls. The finding that output-side geometry contributes almost nothing is a useful, non-obvious result. The selective deployment on vulnerable layers is a reasonable systems contribution, and the 7.6% end-to-end speedup on 64 GH200s is real, though modest relative to the 10.8% from plain Euclidean FP8. Credit for the error-feedback integration in transformed coordinates — that is a clean design choice, not a bolt-on. The paper also honestly reports that GIFT is slower than Euclidean FP8, which is the right framing. The stress-test concern about downstream evaluation lands. Table III is single-seed, 14 tasks, no error bars, and the 7-vs-5 win count over FP32 masks real per-task regressions (CB-ACC at 600M: GIFT 0.30 vs Euclidean 0.41). The 7-vs-5 margin on 14 binary comparisons without significance testing could easily be noise. The paper acknowledges this in §VI.G, but the central practical claim — better downstream preservation — rests on this fragile evidence. The free parameters (rank 32, 13 layers, K=50) are tuned on the same models used for evaluation, which is a mild concern but not disqualifying since the round-trip analysis partially validates the choices independently of downstream metrics. The reader's concern about stale curvature at K=50 is legitimate but speculative — there is no evidence either way here, and it is a testable question for revision. The models are small (300M/600M), which limits generalization claims, but the mechanistic argument does not obviously break at scale. This paper is for distributed systems researchers working on communication compression for LLM pretraining. The coordinate-change principle is worth taking seriously even at this scale. It deserves a serious referee who can push for multi-seed downstream evaluation, sensitivity analysis on the free parameters, and ideally one larger model run. The mechanistic work earns the review; the downstream evidence does not yet close the loop.","headline":"K-FAC whitening repurposed for FP8 gradient communication coordinates — mechanistically clean, downstream evidence thin","tokens_in":18444,"tokens_out":592,"would_cite":false,"duration_ms":85754,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Change coordinates, not precision: whitening gradients before FP8","keywords":[],"falsifier":"Measure the FP8 round-trip reconstruction error of GIFT versus the Euclidean baseline as a function of steps since the last factor-A refresh. If at step 49 (just before refresh) the error matches or exceeds the Euclidean baseline, the stale-factor assumption is violated and the method's benefit is contingent on a refresh interval that may not generalize.","tokens_in":17742,"feed_emoji":"🧭","tokens_out":1080,"duration_ms":155383,"temperature":0.7,"pith_summary":"This paper argues that the dominant source of error in low-precision gradient communication for LLM pretraining is not the number of bits per se, but a mismatch between the anisotropic shape of the gradient distribution and the axis-aligned grid of the quantization format. When gradients are long and stretched in certain directions (as they typically are in Euclidean parameter coordinates), a single FP8 scale factor must stretch to cover the largest direction, leaving smaller directions severely distorted. The authors propose GIFT, which transforms gradients into a near-isotropic coordinate system using a K-FAC-style local curvature metric before quantizing and communicating them, then maps the result back. The key finding is that this coordinate change, applied selectively to the most vulnerable layers and approximated cheaply via a rank-32 input-side factor, makes FP8 communication substantially more faithful to the FP32 reference than direct Euclidean FP8 — improving downstream task preservation while retaining most of the communication speedup.","feed_headline":"Change coordinates, not bits: whitening cuts FP8 gradient error 67%","feed_subtitle":"Transforming gradients into curvature-aware coordinates before FP8 communication preserves downstream model quality while retaining most of ","key_machinery":"The K-FAC block-diagonal approximation to the Fisher information matrix, which decomposes each layer's local curvature into an input-side factor A and an output-side factor G. GIFT uses only the input-side factor A, further approximated at rank 32, to whiten the gradient before FP8 quantization and communication, then applies the inverse transform afterward. The optimizer itself is never modified — the geometry is used purely as a temporary communication coordinate.","core_discovery":"The paper's central discovery is that nearly all the fidelity benefit of full two-sided K-FAC whitening for FP8 gradient communication is captured by the input-side factor alone, and that this factor can be further compressed to rank 32 and deployed on a small subset of layers without meaningful degradation. In a one-step FP8 round-trip test, the input-side transform alone reduced relative reconstruction error by roughly 67% over the Euclidean baseline, matching the full two-sided K-FAC transform almost exactly. This means the practical geometry needed to protect low-precision communication is far lower-dimensional and cheaper to compute than the full second-order curvature would suggest.","pith_inferences":["If the input-side factor A is updated every 50 steps but the true curvature rotates faster (e.g., during sharp loss landscape transitions or learning-rate warmup phases), the whitening transform could become stale and actually increase distortion relative to Euclidean FP8. A test would be to measure reconstruction error as a function of steps-since-last-refresh during high-curvature training phase","The vulnerability profiling identifies fc2 layers as most sensitive in one model profile, but the generality of this finding to other architectures (e.g., non-Llama models, mixture-of-experts) is untested. If the vulnerable-layer set is architecture-specific, the profiling step becomes a mandatory per-model cost rather than a one-time setup.","The 7.6% end-to-end speedup is measured at 64 GPUs; if the geometry-transform overhead is compute-bound while communication savings are bandwidth-bound, the net speedup could either grow (communication dominates at larger scale) or shrink (transform overhead grows with model size) in ways not extrapolated by the current data."],"forward_implications":["If coordinate choice matters more than raw bit-width, then future FP4 gradient communication (when hardware supports it) should benefit from the same whitening principle, since the anisotropy mismatch only worsens at lower bit-widths.","The finding that input-side geometry alone captures nearly all the benefit suggests that output-side curvature, which is more expensive to maintain, may be dispensable for communication purposes even if it matters for optimization.","Selective deployment based on numerical vulnerability profiling implies that the cost of geometry-aware communication can be tuned per model and per training recipe, rather than requiring a fixed universal overhead.","The gap between validation-loss similarity and downstream-task divergence suggests that communication fidelity affects learned representations in ways that standard training-loss monitoring does not capture, raising questions about what loss curves actually measure."],"fun_headline_variants":["Input-side K-FAC factor alone matches full whitening for FP8 gradients","One K-FAC factor, not two, is enough for faithful FP8 gradient comm","Rank-32 single-side whitening captures full geometry for FP8 comm","Compress curvature to rank 32: FP8 gradients stay nearly as faithful","Half of K-FAC whitening cuts FP8 gradient error 67% as well as full"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The input-side K-FAC factor, approximated at rank 32 and refreshed every 50 training steps, is assumed to capture the dominant anisotropic geometry of the gradient throughout training. If the curvature changes faster than the 50-step refresh interval or if rank 32 is insufficient for larger models, the whitening becomes stale and the fidelity benefit degrades.","fun_headline_variants_meta":{"raw":{"variants":["Input-side K-FAC factor alone matches full whitening for FP8 gradients","One K-FAC factor, not two, is enough for faithful FP8 gradient comm","Rank-32 single-side whitening captures full geometry for FP8 comm","Compress curvature to rank 32: FP8 gradients stay nearly as faithful","Half of K-FAC whitening cuts FP8 gradient error 67% as well as full","Input-side curvature alone reduces FP8 gradient error by 67%","Practical geometry for FP8 grads: one factor, rank 32, sparse layers","Cheap geometry: one K-FAC factor at rank 32 protects FP8 communication","Most FP8 gradient protection comes from one low-rank curvature factor","Only input-side curvature matters for FP8: rank 32 suffices","Whiten one side, not two: rank-32 input factor preserves FP8 fidelity","Gradient geometry for FP8 is low-dimensional: one factor, rank 32"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":2320,"prompt_tokens":556,"completion_tokens":1764,"prompt_tokens_details":null},"tokens_in":556,"tokens_out":1764,"duration_ms":59956,"temperature":1.0,"reasoning_tokens":1507,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T09:00:43.230850+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"Measure the FP8 round-trip reconstruction error of GIFT versus the Euclidean baseline as a function of steps since the last factor-A refresh. If at step 49 (just before refresh) the error matches or exceeds the Euclidean baseline, the stale-factor assumption is violated and the method's benefit is contingent on a refresh interval that may not generalize.","supporting_citations":[],"review_version":1}