{"id":"512a3c8e-82fe-4c84-9ad1-7fc466b78b8d","arxiv_id":"2412.06858","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Noise Perturbation Fine-tuning (NPFT) reduces the quantization sensitivity of outlier weights via LoRA fine-tuning with random perturbations, letting standard RTN quantization match GPTQ on LLaMA2-7B-4bit without mixed-precision outlier retention.","lead":"This paper introduces a fine-tuning step that adds random noise to the most quantization-sensitive weights of a large language model, making those weights more robust so the whole model can be safely quantized to 4 bits without keeping special high-precision outliers. The result is a simpler, fully uniform quantized model that runs faster on GPUs while matching the perplexity of more complex mixed-precision schemes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No-noise LoRA control is missing, so the causal claim that perturbation-based Hessian regularization drives NPFT's gains is unsupported.","rationale":"The reader's verdict is CONDITIONAL, and I agree that the paper has concrete numerical inconsistencies and a theoretically shaky identification of Hessian with Fisher information. However, the single most load-bearing concern is more basic: the experimental design never isolates the effect of the noise perturbation itself. The objective in Eq. 7 includes a clean LoRA loss term, so the model is always fine-tuned on calibration data. Standard LoRA fine-tuning alone is known to adapt models to a calibration distribution, and this adaptation can improve post-quantization perplexity regardless of the noise mechanism. Without a z_i=0 control, the paper cannot distinguish between 'NPFT works because perturbation regularizes outlier Hessian trace' and 'NPFT works because LoRA fine-tuning on the calibration data helps.' This is a causal attribution problem, not just a distribution-mismatch problem. The proposed concrete test would settle the question directly by comparing RTN+NPFT against LoRA-only fine-tuning with identical hyperparameters. If they match, the central claim about Hessian regularization is not supported; if they differ substantially, the perturbation is doing real work. Because the reader already required additional evidence before relying on the method, I keep the verdict as CONDITIONAL (UNCHANGED), but I add this control as a necessary condition.","tokens_in":12250,"tokens_out":4635,"duration_ms":48275,"concrete_test":"Re-run the NPFT pipeline with the perturbation disabled: set z_i=0 in Eq. 7 (equivalently minimize (1+β)L(W+U^T V)) using identical outlier selection, LoRA rank, β=0.5, epochs, LR, and C4 calibration data, for OPT-1.3B-4bits and LLaMA2-7B-4bits. Quantize with RTN and report C4/WikiText PPL, plus the outlier Fisher/Hessian-trace sensitivity before and after. If the no-noise control matches RTN+NPFT (within ~0.1 PPL), the noise perturbation is not the active ingredient; if it is close to the RTN baseline and shows no sensitivity reduction, the causal claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central causal claim—that adding random weight perturbations to outliers during LoRA fine-tuning reduces outlier sensitivity and thereby improves quantization—is not established because the paper never ablates the noise. The training objective (Eq. 7), min_{U,V} [L(W+z_i+U^T V)+βL(W+U^T V)], contains a clean-loss term βL(W+U^T V). Setting z_i=0 leaves standard LoRA fine-tuning on the same calibration data (with loss scaled by 1+β). Tables 1-2 compare RTN+NPFT to RTN, GPTQ, etc., but not to 'LoRA fine-tuning without noise.' Fig. 4 sweeps perturbation ratio but starts at 0.05%, not 0. Table 7 varies noise distribution and sampling but always includes noise. Consequently, the reported PPL gains could come entirely from LoRA adaptation/overfitting to the calibration data, not from the perturbation-based implicit Hessian regularization. This is load-bearing because the paper's contribution and theoretical motivation (Eqs. 4-6) are specifically about noise perturbation; if a no-noise LoRA control achieves the same quantized PPL, the method reduces to LoRA fine-tuning and the Hessian-trace mechanism is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Noise Perturbation Fine-tuning (NPFT), a parameter-efficient fine-tuning method that adds random perturbations to a small set of outlier weights (identified by Fisher sensitivity) while training LoRA adapters. The stated goal is to reduce the Hessian trace with respect to those outliers, making the model more robust to post-training quantization so that uniform quantization such as RTN can approach the performance of mixed-precision methods. The authors report perplexity gains on OPT and LLaMA models for RTN, GPTQ, and sqLLM quantizers, including a headline claim that RTN+NPFT matches GPTQ on LLaMA2-7B-4bit, plus latency/memory efficiency comparisons on a 4090 GPU.","tokens_in":12458,"tokens_out":5226,"duration_ms":44078,"significance":"If the mechanism were fully established, NPFT would be a practically valuable contribution: it promises a way to avoid hardware-unfriendly mixed-precision outlier retention while improving quantized model quality, with modest training cost. The paper also provides useful engineering data: one-shot fine-tuning that benefits multiple bit-widths, latency/memory profiling against a QAT baseline, and ablations over noise distribution, layer choice, and calibration data. The main empirical observation, that NPFT improves quantized perplexity across models and quantizers, is credible as reported. However, the central causal claim—that the noise perturbation specifically drives the improvement via Hessian-trace regularization—is not supported by the current experimental design, because no no-noise LoRA fine-tuning control is presented.","major_comments":[{"comment":"The paper never ablates the noise perturbation itself. The training objective in Eq. (7) contains a clean-loss term βL(W+UᵀV); setting z_i=0 reduces the method to standard LoRA fine-tuning on the same calibration data. Tables 1–2 compare RTN+NPFT against RTN, GPTQ, etc., but not against “LoRA fine-tuning without noise,” and Fig. 4 sweeps the perturbation ratio starting at 0.05% rather than 0. Consequently, the reported PPL gains could in principle come entirely from LoRA adaptation to the calibration data, with the noise contributing little or nothing. This is load-bearing because the paper's title, abstract, and theoretical motivation (Eqs. 4–6) attribute the gains specifically to perturbation-based implicit Hessian regularization. The authors should add a no-noise LoRA control (z_i=0, same hyperparameters) to Tables 1–2 and include a 0% perturbation point in Fig. 4; if the no-noise control matches the NPFT results, the mechanism claim would need to be substantially revised.","section":"§3.3, Eq. (7); Tables 1–2; Fig. 4"},{"comment":"The noise distribution used in the implementation does not match the theoretical derivation. Eq. (6) assumes z is uniform on [−Δ/2, Δ/2], with covariance (Δ²/12)I, so that the expected loss under perturbation is proportional to Tr(H). Algorithm 1, however, samples per-channel noise from the channel's weight range (max−min), centered by subtracting the mean; the noise magnitude is therefore determined by the channel's data range, not by the quantization bin width Δ of the target bit-width. This breaks the proportionality Tr(H) ∝ E_z[L(W+z)−L(W)] because Cov(z) is no longer a constant multiple of the identity (different channels have different scales, and the scale is unrelated to Δ). If the intent is to mimic RTN rounding error, the perturbation should be tied to Δ or to the actual rounding residuals; as written, the link between the synthetic noise and the quantization error that motivates Eqs. (5)–(6) is unsubstantiated.","section":"§3.2, Eqs. (5)–(6); §3.3, Algorithm 1"},{"comment":"The 'theoretical insight verification' is circular. Fig. 3 reports that after NPFT, the Fisher sensitivity F decreases for the selected outlier positions. But F is exactly the statistic used in Algorithm 1 to choose the outliers in the first place. Showing that a quantity decreases after fine-tuning when it was used as a selection criterion does not independently confirm that the mechanism is Hessian-trace reduction; it could simply reflect regression to the mean or the effect of LoRA fine-tuning on the Fisher diagonal. A direct estimate of the Hessian trace (e.g., using the Hutchinson estimator on a subset of weights) before and after NPFT, or a per-layer breakdown of quantized PPL changes, would be more conclusive.","section":"§4.3, Fig. 3"},{"comment":"The stated improvement for OPT-1.3B-4bit is inconsistent. The introduction claims 'NPFT helps RTN achieve a 3.69 perplexity improvement on OPT-1.3B-4bits,' but Table 1 reports RTN at 24.68 and RTN+NPFT at 21.74 on C4, which is a 2.94 PPL improvement. The Figure 1 caption likewise says 'over 2.9/0.9,' not 3.69. Please reconcile the numbers; if 3.69 refers to a different configuration or dataset, state that clearly. This matters because the large claimed improvement is highlighted in the contribution summary.","section":"Introduction (contribution list) and Table 1"}],"minor_comments":[{"comment":"The text states that NPFT 'enabl[es] us to perform full-parameter fine-tuning of LLaMA2-7B on a single V100 GPU,' but the method uses LoRA adapters, which is parameter-efficient fine-tuning, not full-parameter fine-tuning. Please correct this phrasing or clarify what is meant.","section":"§4.4, training efficiency discussion"},{"comment":"The text in §4.4 refers to 'shown in Tab. 11' when the training-time comparison appears as Table 4, and an unnumbered Table 11 is placed in the appendix after Table 10. Please renumber the tables consistently and fix the in-text references.","section":"Table numbering"},{"comment":"The pseudocode samples noise as torch.rand_like(W[:, col_idx]) * (max − min) + min, followed by subtracting the mean. This produces a uniform distribution over the channel's data range, not over [−Δ/2, Δ/2] as claimed in §3.2. The pseudocode should be aligned with the theoretical derivation, or the derivation should be revised to cover the actual noise distribution.","section":"Algorithm 1, line 11"},{"comment":"Equation (3) writes H ≈ F = ggᵀ = ||∂L/∂w_{i,j}||²₂. The last equality conflates a full outer-product (or its expectation) with a single squared gradient component; the notation should distinguish the Fisher matrix from its diagonal approximation used for sensitivity scoring.","section":"Eq. (3)"},{"comment":"The text claims 'a 10% reduction in inference latency' from Table 3, but for LLaMA2-7B-4bits the latency is identical (4.80 s) for sqLLM and sqLLM+NPFT; the 10% figure comes only from OPT-1.3B. Please qualify the claim accordingly.","section":"Table 3 and efficiency claim"},{"comment":"Reference [11] attributes GPT-3 to Floridi and Chiriatti rather than to the original Brown et al. paper; please use the canonical citation. Also, the 'Preprint at arXiv' entry for the LLaMA paper is incompletely formatted.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The missing no-noise LoRA control is the key issue; the reviewer's skeptic note correctly identifies it. The empirical PPL improvements are plausible, but the paper's central mechanistic claim (that noise perturbation drives the gains) cannot be evaluated without that control. This is fixable within the manuscript's scope by adding a straightforward ablation. The noise-distribution mismatch between §3.2 and Algorithm 1 is a second substantive issue that should be addressed in revision. I also recommend the editors double-check the reported 3.69 vs 2.94 improvement because such discrepancies in the headline numbers erode confidence in the rest of the evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read through NPFT. The headline result—RTN+NPFT matching GPTQ on LLaMA2-7B-4bit—is genuinely interesting, and the latency win from dropping mixed-precision outlier storage is real. The paper is also honest about being a combination of known pieces: randomized trace estimation (Avron-Toledo), Hessian-aware sensitivity (HAWQ/HERO), Fisher-based outlier selection (SqueezeLLM), and LoRA. The new bit is using LoRA fine-tuning with weight perturbations to reduce outlier sensitivity before quantization, and that combination is worth testing.\n\nWhat the paper does well: clear experiments across OPT and LLaMA, both uniform and non-uniform quantizers, multiple calibration datasets, and a reasonable efficiency comparison. The consistency of the gains is encouraging.\n\nNow the soft spots, in rough order of importance.\n\nFirst, the missing control. The objective in Eq. 7 has a clean-loss term βL(W+U^TV). Setting z_i=0 gives standard LoRA fine-tuning on the same calibration data. The paper never runs that control. Fig. 4 sweeps perturbation ratio but starts at 0.05%, not 0; Table 7 varies noise distribution but always includes noise. So we don't know whether the PPL gains come from the perturbation-based Hessian regularization or just from LoRA adaptation to the calibration set. That's load-bearing because the paper's whole contribution is the noise mechanism. This needs to be fixed with a z=0 ablation, plus ideally an ablation that perturbs random weights rather than Fisher-selected outliers.\n\nSecond, there is a concrete numerical discrepancy: the text and abstract claim a 3.69 PPL improvement on OPT-1.3B-4bit, but Table 1 shows 24.68 to 21.74, a 2.94 improvement. The WikiText numbers line up (11.89), so it looks like a typo, but it needs correcting.\n\nThird, Section 4.3 claims to verify the theoretical insight by showing Fisher sensitivity drops after NPFT. That's measuring the same statistic used to identify outliers, so it's not an independent check on Hessian trace reduction. The authors should either measure something closer to the actual quantization error or at least acknowledge the circularity.\n\nFourth, the paper identifies outliers as the top γ% Fisher positions, but Algorithm 1 adds noise to entire channels containing outliers, not just to outliers. The text says 'add random weight perturbations on the outliers,' but the implementation perturbs all weights in selected output channels. That's fine if stated clearly, but it muddies the mechanism.\n\nAlso, no code or error bars are provided, which makes the quantitative claims harder to evaluate. The reference list looks fine; the relevant prior work is cited. The derivation from Eq. 4 to Eq. 6 is standard. The Fisher-as-Hessian approximation is a known heuristic, not a new error.\n\nBottom line: this is a plausible and useful recipe, but the causal claim is unsupported at present. A no-noise LoRA control is the single most important addition. If you're going to build on this, wait for the revised version. I'd send it to review—it's the kind of empirical paper where a serious referee can push for the missing ablation and the paper could come out solid.","headline":"Useful empirical recipe for quantizing LLMs without outlier retention, but the paper never shows that the noise perturbation itself is what does the work.","tokens_in":13012,"tokens_out":3538,"would_cite":false,"duration_ms":32337,"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":"By adding per-channel random noise to outlier weights during a short LoRA fine-tuning, NPFT shrinks their Hessian sensitivity, letting plain round-to-nearest quantization match GPTQ at 4 bits on LLaMA2-7B with no outliers preserved in…","keywords":["LLM quantization","outlier weights","Hessian trace","noise perturbation fine-tuning","LoRA","post-training quantization","round-to-nearest quantization","perplexity"],"falsifier":"On a model small enough to form its Hessian (for example OPT-125M), measure the Hessian trace restricted to the outlier channels before and after NPFT; the mechanism predicts a substantial drop. A second, quantizer-level test: fine-tune two copies of one model, one with NPFT's synthetic uniform noise and one with perturbations drawn from the actual RTN or GPTQ rounding residuals on the calibration set, then compare 4-bit perplexity; if the residual-matched copy does not match or beat the synthetic-noise copy, the surrogate assumption, rather than Hessian-trace reduction itself, is what carried the result.","tokens_in":12008,"feed_emoji":"🎯","tokens_out":16554,"duration_ms":136331,"temperature":0.7,"pith_summary":"Large language models hide a small set of outlier weights that, when rounded to low precision, wreck the model's predictions. Existing quantizers protect these outliers by keeping them in floating point, forcing a mixed-precision format that slows GPU inference. This paper claims the outliers can instead be tamed: adding per-channel random noise to their locations during a short LoRA fine-tuning implicitly shrinks the loss Hessian trace at those weights, making them robust to rounding. After this tuning, the simplest round-to-nearest quantizer matches the more elaborate GPTQ method at 4 bits on LLaMA2-7B (7.42 perplexity on C4), and quantized models no longer need FP16 outliers, cutting inference latency by roughly a tenth. The payoff is that hardware-friendly single-precision quantization can replace mixed-precision schemes without giving up quality.","feed_headline":"Noise tuning lets plain rounding match GPTQ on 4-bit LLMs","feed_subtitle":"One hour of LoRA tuning tames outlier weights, so no FP16 extras are needed and inference runs about 10% faster.","key_machinery":"The carrying object is the identity $\\operatorname{Tr}(H) = \\mathbb{E}_z[z^T H z]$ for random vectors $z$ with zero mean and identity covariance, joined with the second-order Taylor expansion of the loss. This converts the expensive goal of shrinking the Hessian trace at outlier weights into a cheap one: minimize the expected loss $\\mathbb{E}_z L(W+z)$ under random perturbations. The paper instantiates that with per-channel zero-mean uniform noise, scaled to each outlier-containing channel's value range (mimicking the $\\pm\\Delta/2$ error bound of round-to-nearest quantization), applied only to the top-$\\gamma\\%$ of weights ranked by the Fisher Information Matrix $F = \\|\\partial L/\\partial w_{i,j}\\|_2^2$. LoRA adapters carry the optimization through the objective $\\min_{U,V}\\big[L(W+z_i+U^T V)+\\beta L(W+U^T V)\\big]$ and merge back afterward, keeping the whole procedure to about an hour on a 7B model: the regularization is implicit, and no Hessian is ever formed.","core_discovery":"NPFT's central claim is that the sensitivity of outlier weights can be reduced directly, so that special handling is unnecessary. For a converged model, Taylor expansion says the loss increase under a weight perturbation $z$ is approximately $\\tfrac{1}{2} z^T H z$, so its expectation over zero-mean, identity-covariance perturbations is proportional to the Hessian trace; since round-to-nearest error is bounded by half a quantization bin, the paper samples zero-mean uniform noise on each channel that contains outliers as a stand-in for that error. Fine-tuning with LoRA to minimize the expected loss under this noise, while a weighted clean-loss term guards base performance, then acts as cheap Hessian regularization, avoiding any higher-order gradient computation. The paper's headline result is that with NPFT the simple RTN quantizer reaches 7.42 PPL on C4 for LLaMA2-7B-4bits, identical to GPTQ, and that NPFT improves both uniform and non-uniform quantizers across OPT and LLaMA models, with sqLLM+NPFT matching the mixed-precision sqLLM at lower latency.","pith_inferences":["Because any zero-mean perturbation whose covariance is proportional to the identity yields the same trace identity, the specific uniform distribution is probably not load-bearing; the paper's own ablation with uniform, Gaussian, and Laplace noise points the same way, and a natural extension is to draw perturbations from the target quantizer's actual rounding residuals instead of from a channel's d","The noise is synthesized per channel from that channel's min-max range, so channels with wider ranges receive bigger perturbations and should dominate the training signal; a testable prediction is that outlier sensitivity drops most in exactly those channels.","If the mechanism is genuinely Hessian-trace reduction rather than generic LoRA adaptation, then on a small model where the Hessian is computable, the measured trace on outlier channels should fall after NPFT by an amount predictable from the perturbed-loss objective, and a LoRA-only control without noise should show little or no quantization gain; neither check appears in the paper.","NPFT is complementary to quantizer design: its benefit is largest for the crudest quantizer (RTN) and smallest for the already sensitivity-aware sqLLM, so stacking it with activation-aware scaling or non-uniform schemes may extend gains into the 3-bit and sub-3-bit regimes that this paper does not reach (its RTN+NPFT still diverges at 3 bits on OPT)."],"forward_implications":["The simplest quantizer, RTN, becomes competitive with GPTQ after NPFT on LLaMA2-7B-4bits (7.42 vs. 7.42 PPL on C4), so the costliest parts of post-training quantization may be unnecessary once the weights are noise-regularized.","Outlier weights are quantized at the same bit-width as all other weights, eliminating the FP16 sparse matrices whose mixed-precision format slows GPU kernels and wastes memory bandwidth.","The gains transfer across quantizer families (uniform RTN and GPTQ, non-uniform sqLLM) and model families (OPT, LLaMA), including unseen datasets such as WikiText and PTB and reasoning tasks such as PIQA, ARC, and Storycloze, so the tuning is not tailored to one quantizer.","A single NPFT pass serves multiple bit-widths at once and uses roughly a quarter of EfficientQAT's training time on LLaMA2-7B, lowering the adoption barrier for retraining-based quantization.","With the sqLLM quantizer, NPFT's single-precision 4-bit model matches the mixed-precision baseline's perplexity while cutting generation latency by about 10% and slightly reducing GPU memory."],"supporting_citations":[{"why":"Supplies the randomized trace-estimation identity Tr(H) = E[z^T H z] that lets Hessian regularization be replaced by noise perturbation.","marker":"[1]"},{"why":"The quantization-aware-training baseline whose training time and memory usage NPFT is compared against.","marker":"[2]"},{"why":"Documents the channel-wise concentration of outlier weights that motivates per-channel noise sampling; also a mixed-precision baseline NPFT replaces.","marker":"[6]"},{"why":"Motivates using Hessian eigenvalues and trace as weight-sensitivity criteria for quantization decisions.","marker":"[8]"},{"why":"The post-training quantizer whose LLaMA2-7B 4-bit C4 result (7.42 PPL) RTN+NPFT is claimed to match.","marker":"[12]"},{"why":"LoRA adapters are the parameter-efficient machinery that keeps the fine-tuning short.","marker":"[13]"},{"why":"Provides the Fisher-based outlier identification adopted by NPFT, plus the sqLLM quantizer and GPU kernels used in latency and memory profiling.","marker":"[15]"},{"why":"Activation-distribution-based outlier handling baseline that appears in the main comparison tables.","marker":"[16]"},{"why":"Proved that regularizing Hessian eigenvalues improves quantization robustness, the conceptual precursor NPFT makes tractable for LLMs.","marker":"[25]"}],"fun_headline_variants":["Noise fine-tuning tames outlier weights, making simple RTN match GPTQ","Perturb outliers during LoRA to boost LLM quantization without mixed precision","4-bit LLMs: noise perturbation fine-tuning beats special outlier handling","Tame sensitive weights with noise to get GPTQ-level 4-bit accuracy","RTN becomes competitive with GPTQ after noise-perturbation fine-tuning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on treating random zero-mean uniform noise spread over outlier channels as a faithful stand-in for the real, deterministic rounding errors the quantizer will later make; if those two distributions disagree, making the model robust to the noise will not make it robust to quantization.","fun_headline_variants_meta":{"raw":{"variants":["Noise fine-tuning tames outlier weights, making simple RTN match GPTQ","Perturb outliers during LoRA to boost LLM quantization without mixed precision","4-bit LLMs: noise perturbation fine-tuning beats special outlier handling","Tame sensitive weights with noise to get GPTQ-level 4-bit accuracy","RTN becomes competitive with GPTQ after noise-perturbation fine-tuning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000593,"raw_usage":{"total_tokens":2803,"prompt_tokens":993,"completion_tokens":1810,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":1710}},"tokens_in":609,"tokens_out":1810,"duration_ms":13142,"temperature":1.0,"reasoning_tokens":1710,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:02:56.949466+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a model small enough to form its Hessian (for example OPT-125M), measure the Hessian trace restricted to the outlier channels before and after NPFT; the mechanism predicts a substantial drop. A second, quantizer-level test: fine-tune two copies of one model, one with NPFT's synthetic uniform noise and one with perturbations drawn from the actual RTN or GPTQ rounding residuals on the calibration set, then compare 4-bit perplexity; if the residual-matched copy does not match or beat the synthetic-noise copy, the surrogate assumption, rather than Hessian-trace reduction itself, is what carried the result.","supporting_citations":[{"cited_title":"Randomized algorithms for estimating the trace of an implicit symmetric positive semi-definite matrix","cited_arxiv_id":null,"evidence_quote":"Supplies the randomized trace-estimation identity Tr(H) = E[z^T H z] that lets Hessian regularization be replaced by noise perturbation."},{"cited_title":"Hawq: Hessian aware quantization of neural networks with mixed-precision","cited_arxiv_id":null,"evidence_quote":"Motivates using Hessian eigenvalues and trace as weight-sensitivity criteria for quantization decisions."},{"cited_title":"Awq: Activation-aware weight quantization for on-device llm compression and acceleration","cited_arxiv_id":null,"evidence_quote":"Activation-distribution-based outlier handling baseline that appears in the main comparison tables."},{"cited_title":"Hero: Hessian-enhanced robust optimization for unifying and improving generalization and quantization performance","cited_arxiv_id":null,"evidence_quote":"Proved that regularizing Hessian eigenvalues improves quantization robustness, the conceptual precursor NPFT makes tractable for LLMs."}],"review_version":1}