Pith. sign in

REVIEW 2 major objections 4 minor 117 references

Quantizing a weight-tied reasoner at 4-bit per-tensor turns a small per-step rounding error into a coherent drift that destroys exact solutions; switching to per-block activation scaling restores near-full-precision accuracy.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 09:54 UTC pith:NR537XOW

load-bearing objection The drift diagnosis and MXInt4 fix are solid and well-triangulated, but the core mechanism claim skips a per-tensor affine (zero-point) control that could reframe the whole story. the 2 major comments →

arxiv 2607.16237 v1 pith:NR537XOW submitted 2026-06-25 cs.LG cs.AI

Quantizing Recursive Reasoning Models

Thorir Mar Ingolfsson , Wajeeha Tahir , Anna Tegon , Lionnus Kesting , Gamze \.Islamo\u{g}lu , Luca Benini This is my paper
classification cs.LG cs.AI
keywords recursive reasoning modelsquantizationactivation scaling granularityMXInt4post-training quantizationtrajectory driftexact accuracyweight-tied transitions
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper studies what happens when a recursive reasoning model—a small weight-tied block applied many times to refine a puzzle solution—is quantized to low precision. It shows that 8-bit quantization behaves like a small perturbation that the recursion damps, but per-tensor 4-bit quantization injects a systematic bias that accumulates on every reuse, dropping exact Sudoku accuracy from 84.1% to 0.0%. The central claim is that the collapse is caused by the granularity of activation scaling, not by bit-width or integer-versus-float format: switching from one shared scale for all activations to a local scale per block of 32 values (MXInt4) restores near-full-precision accuracy on Sudoku, Maze, and ARC-AGI. The paper also shows that deeper recursions are more sensitive, and that neither test-time noise nor standard outlier corrections fix the bias. The result matters because tiny recursive reasoners are attractive for low-power deployment, but only if they survive low-bit inference.

Core claim

The discovery is a failure mode with a simple fix. In a weight-tied recursion, the same quantized transition is applied hundreds of times, so the quantizer's one-step error becomes a trajectory-level drift: a nonzero mean perturbation pushes the reasoning state away from the full-precision solution path at every step. Per-tensor Int4 and FP4 both collapse exact accuracy (0.0% and 0.6% on Sudoku), while per-block MXInt4 and MXFP4 recover to about 80.1% and 72.7%; the same pattern transfers from MLP-based Sudoku to attention-based Maze and to ARC-AGI pass@2. The paper identifies activation-scaling granularity as the operative control variable, with controls ruling out calibration freshness, ca

What carries the argument

The central object is the reused recursive transition operator applied to a latent state and an answer embedding, analyzed as a perturbed dynamical system. Quantization turns it into the full-precision transition plus an error term; the key distinction is whether the error's mean component is nonzero and aligned across steps—a coherent bias that accumulates into a fixed-point displacement—or zero-mean noise that averages away. The proposed fix is MXInt4, a blockwise integer activation format that gives each block of 32 activations its own power-of-two shared scale, shrinking both the error magnitude and its systematic component so the quantized trajectory contracts back toward the full-preci

Load-bearing premise

The whole mechanism story assumes the full-precision recursion pulls nearby reasoning states together, and that this pulling force stays stronger than the quantizer's push-apart effect near the solution region; that push-apart effect is never measured, so the drift explanation and the granularity-versus-bit-width causal claim rest on observed behavior rather than a verified inequality.

What would settle it

Measure per-step latent divergence (distance between quantized and full-precision internal states) on the released Sudoku checkpoint while sweeping the number of inner recursion cycles from 1 to 8. The accumulation claim predicts monotonic growth for per-tensor Int4 and contraction for per-block MXInt4; a single run where per-tensor Int4 divergence contracts at high reuse—or where exact accuracy survives—would settle against the mechanism.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Per-tensor 4-bit post-training quantization should be considered unsafe for weight-tied recursive reasoners; per-block activation scaling is the recovery mechanism, not bit-width or number format.
  • Deeper recursions are the most quantization-sensitive: an equilibrium-style model that reuses its block about 288 times collapses to near-zero under per-tensor Int4, while per-block scaling restores both accuracy and depth-scaling behavior.
  • Standard activation-outlier corrections such as SmoothQuant and Hadamard rotation do not rescue per-tensor 4-bit; only local blockwise scaling works at post-training.
  • Test-time stochasticity and trajectory aggregation do not average away the bias, confirming the error is systematic rather than zero-mean.
  • Quantization-aware training is a complementary route: naive Int4 QAT already recovers much of the loss, and learned weight-side step sizes add further gains on MLP-style recursion.
  • On ARC-AGI-1 and ARC-AGI-2, per-tensor Int4 gives 0.0% pass@2 while MXInt4 matches the full-precision point estimate, showing the rule transfers to open-ended reasoning tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Inference: if the granularity mechanism is as general as the paper suggests, blockwise activation scaling should be tested on recursive language-model variants, where long autoregressive reuse could show the same drift.
  • Inference: the contraction-versus-drift framing implies that recursive reasoners could be made intrinsically more quantizable by regularizing the transition toward stronger contraction near the solution manifold—a design principle beyond the paper's post-training fix.
  • Inference: because exact accuracy is all-or-nothing, the Sudoku split between 0.0% exact accuracy and roughly 25% cell accuracy suggests evaluations of quantized reasoners should always report both; a model can look partially correct while producing zero valid solutions.
  • Inference: the same coherent-bias reasoning likely applies to other iterated shared-weight computations, such as diffusion samplers, but the paper's step-independent quantizer is a special case; comparing blockwise activation scaling with timestep-aware quantizers would be a direct test.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper studies post-training quantization of recursive reasoning models — weight-tied blocks applied many times — and reports a sharp, reusable-transition phenomenon: per-tensor 4-bit quantization of activations induces a systematic, accumulating bias that collapses exact accuracy on Sudoku from 84.1% to 0.0%, while per-block scaling (MXInt4) restores near-FP accuracy (80.1%). The authors argue that activation-scaling granularity, not bit-width or number format, controls the contraction-versus-drift boundary, and support this with a format ladder, a granularity sweep, trajectory diagnostics, reuse-dose-response and untied controls, EqR depth scaling, stochasticity controls, and ARC-AGI transfer. They also present a qualitative dynamical-systems model (Proposition 1) in which a biased transition perturbation displaces the fixed point by roughly ε/(1−L). The empirical MXInt4 result is well triangulated, but the causal attribution to granularity has a missing control.

Significance. If the central claim holds, the paper identifies a practically important and previously uncharacterized failure mode of low-bit recursive inference: error accumulation through a reused quantized transition, and a simple fix (blockwise activation scaling) that works across architectures and transfers to ARC-AGI. The paper has notable strengths: the format ladder (Int4/FP4 vs MXInt4/MXFP4) and granularity ladder isolate a monotone effect; the per-block MXInt4 format is deterministic and has no calibration state; the reuse dose-response, matched-capacity untied control, and stochasticity experiments make the accumulation story concrete; and the ARC-AGI transfer provides an out-of-distribution check. However, the abstract and Section 4.2 state a causal claim — 'granularity, not bit-width or number format' — that is not fully established because the per-tensor baselines are all symmetric (zero-point-free) quantizers. The paper also candidly discloses that Proposition 1's regime condition L+δ<1 is inferred rather than measured. These issues are fixable with targeted experiments and wording changes, so the contribution is promising but not yet ready in its current causal form.

major comments (2)
  1. [§4.2/§4.3, Eq. (3)] The paper's central attribution — 'activation-scaling granularity rather than bit-width or number format' (abstract; §4.2) — is not fully supported because all per-tensor 4-bit baselines are symmetric, zero-point-free quantizers. Equation (3) defines only a symmetric integer quantizer, and 'zero-point' never appears in the manuscript. A per-tensor affine quantizer (min/max or learned zero-point) can remove the constant component of E[Q(a)-a] that §3.3 identifies as the driver, while keeping scale granularity per-tensor. The granularity ladder in §4.3 sweeps group size using symmetric float-absmax scales, so it cannot separate group-size effects from symmetric-vs-affine effects; likewise, the dynamic per-tensor control uses dynamic absmax, not affine min/max. The most natural alternative explanation for the per-tensor 4-bit collapse is therefore uncontrolled. Please add a per-tensor affin
  2. [§3.4 / A.11] Proposition 1's predictive content depends on the unmeasured smoothed Lipschitz constant δ and on the assumption that the full-precision transition is locally contractive with L<1. The paper explicitly labels this a qualitative scaffold and states that δ is not estimated, which is good, but the main text repeatedly uses the mechanism with causal language: 'granularity alone controls the contraction-versus-drift boundary' (§4.2), 'exactly as Section 3.3 predicts' (§4.3), and 'explains why...' (§5.3). As it stands, the regime split L+δ<1 is inferred from the very drift that the model is meant to explain, so statements of this strength go beyond the verified content. Please either (a) provide a direct measurement of δ (e.g., from calibration activations via local Lipschitz estimates of the quantized transition) or (b) consistently hedge the mechanism as a hypothesis that is consistent with,
minor comments (4)
  1. [Eq. (3)] The expression 'qB = clip(j aB/sB m, −qmax, qmax)' is missing a floor or rounding symbol in the typeset version; as written it is not standard math notation.
  2. [§4.3] The phrase '∥d∥/∥d∥ ≈ 0.9' appears to contain a typo or undefined notation; presumably this is intended to be a cosine similarity or a ratio of normalized projections.
  3. [§4.2 / Table 1] On Maze, per-tensor Int4 survives (73.5%) while per-tensor FP4 collapses (0.4%), which the text calls 'orthogonal' to the granularity effect. This is an interesting and unexplained reversal; a sentence explaining why the integer per-tensor grid is better matched to Maze activations would help readers interpret the format ladder.
  4. [§4.6 / Table 4] The ARC-AGI-2 numbers are single-seed and each task is worth 6.25%; the text does note this, but the abstract's phrase 'transferring to ARC-AGI' should be read with that coarse granularity in mind. A short caveat in the abstract or a coarser ARC-AGI-2 claim would be appropriate.

Circularity Check

0 steps flagged

No significant circularity: the paper's derivation is empirical and self-contained.

full rationale

The paper's central conclusions are driven by direct measurements and ablations, not by fitted parameters that are then relabeled as predictions. The MXInt4 block scale is a per-block absmax computed online at inference with no calibration state, so the main 'prediction' (per-block scaling avoids the drift) is not built into the quantizer by construction. The granularity ladder sweeps group size while holding other factors fixed, providing an independent empirical axis rather than a self-defined outcome. Proposition 1 is explicitly labeled a conditional, qualitative scaffold rather than a tight bound, and the paper states that delta is not measured and that the L+delta<1 regime split is inferred from observed drift. That is a limitation of the mechanistic story, not an equation that reduces to its own inputs. There are no load-bearing self-citations: the cited HRM/TRM/EqR works are external baselines or checkpoints, and the microscaling formats are cited to the OCP specification and prior non-author work. The unperformed per-tensor affine (zero-point) control is a genuine experimental-validity concern about the causal attribution to granularity, but it is not a circularity: a missing control does not make the derivation equivalent to its inputs. The paper's empirical recoveries (Sudoku, Maze, EqR, ARC-AGI) are out-of-sample and the depth-scaling comparisons are measured, so the central claim retains independent content. Score 0.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The paper introduces no new theoretical entities; MXInt4 is an existing OCP microscaling format. The main load-bearing assumptions are the unproven local-contraction model and the smoothed-Lipschitz treatment of quantizers. No free parameters are fitted to force the central claim; the block size is a design choice, and α is only a control-sweep parameter.

free parameters (2)
  • Block size B=32 for per-block scaling
    Chosen to match the OCP MX format; the granularity ladder also tests 16/32, but the main MXInt4 result depends on this group-size hyperparameter, not on a value fitted to optimize accuracy.
  • SmoothQuant mixing strength α (control sweep) = α ∈ {0.3, 0.5, 0.7, 0.8, 0.85, 0.95}
    Swept in the outlier-control experiments (Table 3); not used in the final method and not fitted to the central claim.
axioms (4)
  • domain assumption The full-precision recursion Fθ is locally contractive with modulus L<1 near the fixed point s*.
    Section 3.4 / A.11: inferred from empirical residual decay (EqR residual 27.5→16.2; Int8 divergence shrinking). Not proven; if false, the displacement bound ε/(1−L) loses its foundation.
  • domain assumption The quantizer perturbation admits a bounded magnitude ε and a smoothed Lipschitz bound δ on a neighbourhood U.
    Equation (4), Section A.11: piecewise-constant quantizers are not literally Lipschitz, so δ is an expected, smoothed sensitivity. δ is never directly estimated.
  • domain assumption The mean transition perturbation µQ(x) along an input's trajectory drives the observable displacement, and zero-mean fluctuations do not.
    Section 3.3 / A.11: supported behaviorally (test-time noise injection fails to rescue Int4; cross-step cosine 0.99), but only layerwise proxies are measured, not the full per-step mean.
  • domain assumption Exact accuracy (all-or-nothing) is the right metric to judge quantization failure for these reasoning tasks.
    A single wrong symbol zeroes a Sudoku solution; the paper uses exact accuracy throughout. Reasonable for the task family, but it is a metric choice that amplifies small errors.

pith-pipeline@v1.3.0-alltime-deepseek · 26679 in / 10636 out tokens · 104108 ms · 2026-08-02T09:54:39.344921+00:00 · methodology

0 comments
read the original abstract

Recursive reasoning models solve hard puzzles by applying compact, weight-tied blocks over many refinement steps. Because these blocks are reused many times, quantizing them creates a unique dynamical problem: the quantization error is incurred at every step. While 8-bit quantization (integer or float) preserves accuracy, moving to a per-tensor 4-bit format causes a systematic bias to accumulate. The ensuing drift catastrophically degrades exact-solution accuracy on Sudoku from 84.1% to 0.0% (only ~25% of cells correct). In this work, we show that this collapse is caused by activation-scaling granularity rather than bit-width or number format. Crucially, moving to per-block scaling completely restores the transition. To implement this, we apply MXInt4, a blockwise integer activation format, to recursive reasoning models. It is competitive with blockwise float formats on our tasks, while keeping integer elements and power-of-two block scales. Finally, recursion depth and reuse modulate quantization sensitivity, with the deepest architecture we test (the EqR equilibrium model) the most sensitive. Yet blockwise scaling overcomes this vulnerability, staying robust across these architectures and transferring to the open-ended ARC-AGI benchmark.

Figures

Figures reproduced from arXiv: 2607.16237 by Anna Tegon, Gamze \.Islamo\u{g}lu, Lionnus Kesting, Luca Benini, Thorir Mar Ingolfsson, Wajeeha Tahir.

Figure 1
Figure 1. Figure 1: Transition contraction versus drift (Sudoku). [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Granularity, rather than integer vs. float, controls stability. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: EqR mechanism: per-block scaling preserves depth use [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Per-step trajectory atlas (full-coverage diagnostic). Exact accuracy, cell accuracy, and latent divergence d z t vs. recursion step for FP, Int8, per-tensor Int4, and per-block MXFP4, on Sudoku (top) and Maze (bottom). Int8 tracks full precision and its divergence contracts; per-tensor Int4 collapses on the deep Sudoku recursion (flat-high divergence, 0% exact) yet survives the shallow Maze; per-block MXFP… view at source ↗
Figure 5
Figure 5. Figure 5: reports the per-block integer (MXInt) sweep on Sudoku referenced in Section 4.4. This includes the MXInt8 gate result (84.1%, matching FP) and the MXInt4 weight-only point (82.5%). Int4 (per-tensor) MXInt4 (per-block) MXInt8 (per-block) FP 0% 20% 40% 60% 80% 100% exact accuracy (Sudoku) 0.0% 80.1% 84.1% 84.1% [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Activation quantization dominates transition bias. [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Supporting evidence for the recursion-structure analysis. (a) On EqR, breadth improves full precision but not Int4. (b) On TRM, test-time stochasticity leaves Int4 accuracy at zero across all recursion steps. The basin panel of [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Test-time stochasticity does not improve Int4 accuracy. [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Activation granularity controls EqR quantization. [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Dose–response to transition reuse. Sweeping inner recursion cycles L changes how many times the quantized block is applied before read-out (Sudoku L ∈ {1, 2, 4, 6, 8}, Maze L ∈ {1, 2, 4, 8}). (a) On Sudoku the FP–Int4 cell-accuracy gap widens monotonically with reuse (accumulation). (b,c) On the shallower Maze recursion Int4 instead self-corrects: exact accuracy climbs toward FP and latent divergence shri… view at source ↗
Figure 11
Figure 11. Figure 11: Matched-capacity untied control. Untying the L-cycle positions (7× less reuse) roughly doubles Int4 cell-accuracy and lowers final latent divergence. Untied Int4 still has near-zero exact accuracy (1.6%), so reuse amplifies the failure of exact solutions without solely causing it. This happens because the reuse that survives untying is still substantial. Per forward pass, the fully tied model applies each… view at source ↗
Figure 12
Figure 12. Figure 12: Reused transition layers dominate activation error. [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

117 extracted references · 2 canonical work pages

  1. [1]

    2025 , note =

    Less is More: Recursive Reasoning with Tiny Networks , author =. 2025 , note =. 2510.04871 , archivePrefix =

    Pith/arXiv arXiv 2025

  2. [2]

    2025 , note =

    Hierarchical Reasoning Model , author =. 2025 , note =. 2506.21734 , archivePrefix =

    Pith/arXiv arXiv 2025

  3. [3]

    2605.20613 , archivePrefix =

    Wang, Guan and Liu, Changling and Wang, Chenyu and Zhou, Cai and Sun, Yuhao and Wu, Yifei and Zhen, Shuai and Scimeca, Luca and Abbasi Yadkori, Yasin , year =. 2605.20613 , archivePrefix =

    Pith/arXiv arXiv

  4. [4]

    2026 , eprint =

    Probabilistic Tiny Recursive Model , author =. 2026 , eprint =

    2026

  5. [5]

    2026 , eprint =

    Generative Recursive Reasoning , author =. 2026 , eprint =

    2026

  6. [6]

    2026 , note =

    Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning , author =. 2026 , note =. 2605.21488 , archivePrefix =

    Pith/arXiv arXiv 2026

  7. [7]

    2026 , eprint =

    Tiny Autoregressive Recursive Models , author =. 2026 , eprint =

    2026

  8. [8]

    Tiny Recursive Models on

    Roye-Azar, Antonio and Vargas-Naranjo, Santiago and Ghai, Dhruv and Balamurugan, Nithin and Amir, Rayan , year =. Tiny Recursive Models on. 2512.11847 , archivePrefix =

    arXiv

  9. [9]

    2025 , eprint =

    Accelerating Training Speed of Tiny Recursive Models with Curriculum Guided Adaptive Recursion , author =. 2025 , eprint =

    2025

  10. [10]

    Tiny Recursive Reasoning with

    Wang, Wenlong and Reid, Fergal , year =. Tiny Recursive Reasoning with. 2602.12078 , archivePrefix =

    arXiv

  11. [11]

    International Conference on Learning Representations (ICLR) , year =

    Universal Transformers , author =. International Conference on Learning Representations (ICLR) , year =

  12. [12]

    Proceedings of the 40th International Conference on Machine Learning , pages =

    Looped Transformers as Programmable Computers , author =. Proceedings of the 40th International Conference on Machine Learning , pages =. 2023 , editor =

    2023

  13. [13]

    Proceedings of the 42nd International Conference on Machine Learning , pages =

    On Expressive Power of Looped Transformers: Theoretical Analysis and Enhancement via Timestep Encoding , author =. Proceedings of the 42nd International Conference on Machine Learning , pages =. 2025 , editor =

    2025

  14. [14]

    2016 , eprint =

    Adaptive Computation Time for Recurrent Neural Networks , author =. 2016 , eprint =

    2016

  15. [15]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Deep Equilibrium Models , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  16. [16]

    2026 , eprint =

    Recursive Models for Long-Horizon Reasoning , author =. 2026 , eprint =

    2026

  17. [17]

    2016 , eprint =

    Recurrent Neural Networks with Limited Numerical Precision , author =. 2016 , eprint =

    2016

  18. [18]

    2016 , eprint =

    Effective Quantization Methods for Recurrent Neural Networks , author =. 2016 , eprint =

    2016

  19. [19]

    2018 , eprint =

    Effective Quantization Approaches for Recurrent Neural Networks , author =. 2018 , eprint =

    2018

  20. [20]

    4-bit Quantization of

    Fasoli, Andrea and Chen, Chia-Yu and Serrano, Mauricio and Sun, Xiao and Wang, Naigang and Venkataramani, Swagath and Saon, George and Cui, Xiaodong and Kingsbury, Brian and Zhang, Wei and T. 4-bit Quantization of. Interspeech , year =

  21. [21]

    Journal of Machine Learning Research , volume =

    Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations , author =. Journal of Machine Learning Research , volume =

  22. [22]

    2018 , eprint =

    Precision Highway for Ultra Low-Precision Quantization , author =. 2018 , eprint =

    2018

  23. [23]

    2023 , eprint =

    Quantized Neural Networks for Low-Precision Accumulation with Guaranteed Overflow Avoidance , author =. 2023 , eprint =

    2023

  24. [24]

    Colbert, Ian and Pappalardo, Alessandro and Petri-Koenig, Jakoba , booktitle =

  25. [25]

    Lan, Zhenzhong and Chen, Mingda and Goodman, Sebastian and Gimpel, Kevin and Sharma, Piyush and Soricut, Radu , booktitle =

  26. [26]

    2026 , eprint =

    Quantization Robustness of Monotone Operator Equilibrium Networks , author =. 2026 , eprint =

    2026

  27. [27]

    International Conference on Learning Representations (ICLR) , year =

    Elucidating the Exposure Bias in Diffusion Models , author =. International Conference on Learning Representations (ICLR) , year =

  28. [28]

    Proceedings of the 40th International Conference on Machine Learning , pages =

    Input Perturbation Reduces Exposure Bias in Diffusion Models , author =. Proceedings of the 40th International Conference on Machine Learning , pages =. 2023 , volume =

    2023

  29. [29]

    International Conference on Learning Representations (ICLR) , year =

    Alleviating Exposure Bias in Diffusion Models through Sampling with Shifted Time Steps , author =. International Conference on Learning Representations (ICLR) , year =

  30. [30]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  31. [31]

    Findings of the Association for Computational Linguistics (ACL) , year =

    Why Exposure Bias Matters: An Imitation Learning Perspective of Error Accumulation in Language Generation , author =. Findings of the Association for Computational Linguistics (ACL) , year =

  32. [32]

    Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages =

    Exposure Bias versus Self-Recovery: Are Distortions Really Incremental for Autoregressive Text Generation? , author =. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages =. 2021 , publisher =. doi:10.18653/v1/2021.emnlp-main.415 , eprint =

    work page doi:10.18653/v1/2021.emnlp-main.415 2021

  33. [33]

    2025 , eprint =

    Beyond Exponential Decay: Rethinking Error Accumulation in Large Language Models , author =. 2025 , eprint =

    2025

  34. [34]

    NeurIPS 2024 Workshop on Audio Imagination: AI-Driven Speech, Music, and Sound Generation , year =

    Continuous Autoregressive Models with Noise Augmentation Avoid Error Accumulation , author =. NeurIPS 2024 Workshop on Audio Imagination: AI-Driven Speech, Music, and Sound Generation , year =

    2024

  35. [35]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Post-training Quantization on Diffusion Models , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  36. [36]

    Li, Xiuyu and Liu, Yijiang and Lian, Long and Yang, Huanrui and Dong, Zhen and Kang, Daniel and Zhang, Shanghang and Keutzer, Kurt , booktitle =

  37. [37]

    2401.04585 , archivePrefix =

    Liu, Xuewen and Li, Zhikai and Xiao, Junrui and Gu, Qingyi , year =. 2401.04585 , archivePrefix =

    Pith/arXiv arXiv

  38. [38]

    2026 , eprint =

    Gradient-Aligned Calibration for Post-Training Quantization of Diffusion Models , author =. 2026 , eprint =

    2026

  39. [39]

    He, Yefei and Liu, Luping and Liu, Jing and Wu, Weijia and Zhou, Hong and Zhuang, Bohan , booktitle =

  40. [40]

    Chu, Huanpeng and Wu, Wei and Zang, Chengjie and Yuan, Kun , booktitle =

  41. [41]

    European Conference on Computer Vision (ECCV) , year =

    Timestep-Aware Correction for Quantized Diffusion Models , author =. European Conference on Computer Vision (ECCV) , year =

  42. [42]

    doi:10.52202/079017-1712 , editor =

    Chen, Yi-Chung and Huang, Zhi-Kai and Chen, Jing-Ren , booktitle =. doi:10.52202/079017-1712 , editor =

    work page doi:10.52202/079017-1712

  43. [43]

    2603.18095 , archivePrefix =

    Ryu, Sooyoung and Salzmann, Mathieu and Javed, Saqib , year =. 2603.18095 , archivePrefix =

    arXiv

  44. [44]

    2025 , eprint =

    Error Propagation Mechanisms and Compensation Strategies for Quantized Diffusion , author =. 2025 , eprint =

    2025

  45. [45]

    2025 , eprint =

    Quantizing Diffusion Models from a Sampling-Aware Perspective , author =. 2025 , eprint =

    2025

  46. [46]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Temporal Dynamic Quantization for Diffusion Models , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  47. [47]

    Li, Yanjing and Xu, Sheng and Cao, Xianbin and Sun, Xiao and Zhang, Baochang , booktitle =

  48. [48]

    Huang, Yushi and Gong, Ruihao and Liu, Jing and Chen, Tianlong and Liu, Xianglong , booktitle =

  49. [49]

    2025 , note =

    Huang, Haocheng and Chen, Jiaxin and Guo, Jinyang and Zhan, Ruiyi and Wang, Yunhong , booktitle =. 2025 , note =

    2025

  50. [50]

    He, Yefei and Liu, Jing and Wu, Weijia and Zhou, Hong and Zhuang, Bohan , booktitle =

  51. [51]

    Pioneering 4-Bit

    Zhao, Maosen and Chen, Pengtao and Yu, Chong and Wen, Yan and Tan, Xudong and Chen, Tao , booktitle =. Pioneering 4-Bit. 2025 , pages =. doi:10.1109/CVPR52734.2025.01690 , note =

    arXiv 2025

  52. [52]

    , year =

    Lewandowski, Basile and Kurz, Simon and Shankar, Aditya and Birke, Robert and Chen, Jian-Jia and Chen, Lydia Y. , year =. 2603.14062 , archivePrefix =

    arXiv

  53. [53]

    Feng, Weilun and Yang, Chuanguang and An, Zhulin and Huang, Libo and Diao, Boyu and Wang, Fei and Xu, Yongjun , booktitle =

  54. [54]

    Wu, Junyi and Wang, Haoxuan and Shang, Yuzhang and Shah, Mubarak and Yan, Yan , booktitle =

  55. [55]

    2509.01624 , archivePrefix =

    Frumkin, Natalia and Marculescu, Diana , year =. 2509.01624 , archivePrefix =

    Pith/arXiv arXiv

  56. [56]

    Conference on Language Modeling (COLM) , year =

    Quantization Hurts Reasoning? An Empirical Study on Quantized Reasoning Models , author =. Conference on Language Modeling (COLM) , year =

  57. [57]

    Quantization Meets Reasoning: Exploring

    Li, Zhen and Su, Yupeng and Yang, Runming and Xie, Congkai and Wang, Zheng and Xie, Zhongwei and Wong, Ngai and Yang, Hongxia , year =. Quantization Meets Reasoning: Exploring. 2501.03035 , archivePrefix =

    Pith/arXiv arXiv

  58. [58]

    2026 , eprint =

    Quantization Meets Reasoning: Exploring and Mitigating Degradation of Low-Bit LLMs in Mathematical Reasoning , author =. 2026 , eprint =

    2026

  59. [59]

    When Reasoning Meets Compression: Understanding the Effects of

    Zhang, Nan and Kwek, Eugene and Zhang, Yusen and Nguyen, Ngoc-Hieu and Mitra, Prasenjit and Zhang, Rui , year =. When Reasoning Meets Compression: Understanding the Effects of. 2504.02010 , archivePrefix =

    arXiv

  60. [60]

    2602.09130 , archivePrefix =

    von Rad, Jonathan and Cao, Yong and Geiger, Andreas , year =. 2602.09130 , archivePrefix =

    Pith/arXiv arXiv

  61. [61]

    2025 , eprint =

    Benchmarking Post-Training Quantization in LLMs: Comprehensive Taxonomy, Unified Evaluation, and Comparative Analysis , author =. 2025 , eprint =

    2025

  62. [62]

    2026 , eprint =

    What Makes Low-Bit Quantization-Aware Training Work for Reasoning LLMs? A Systematic Study , author =. 2026 , eprint =

    2026

  63. [63]

    2602.02581 , archivePrefix =

    Zhang, Nan and Kwek, Eugene and Zhang, Yusen and Pan, Muyu and Wang, Suhang and Mitra, Prasenjit and Zhang, Rui , year =. 2602.02581 , archivePrefix =

    arXiv

  64. [64]

    2605.08755 , archivePrefix =

    Choi, Euntae and Song, Sumin and Yoo, Sungjoo , year =. 2605.08755 , archivePrefix =

    Pith/arXiv arXiv

  65. [65]

    2023 , eprint =

    Microscaling Data Formats for Deep Learning , author =. 2023 , eprint =

    2023

  66. [66]

    2023 , howpublished =

    2023

  67. [67]

    International Conference on Learning Representations (ICLR) , year =

    Is Finer Better? The Limits of Microscaling Formats in Large Language Models , author =. International Conference on Learning Representations (ICLR) , year =

  68. [68]

    2026 , eprint =

    Benchmarking Post-Training Quantization of Large Language Models under Microscaling Floating Point Formats , author =. 2026 , eprint =

    2026

  69. [69]

    Unveiling the Potential of Quantization with

    Chhugani, Jatin and Jeong, Geonhwa and Su, Bor-Yiing and Pan, Yunjie and Yang, Hanmei and Ankit, Aayush and Yu, Jiecao and Deng, Summer and Chen, Yunqing and Satish, Nadathur and Kim, Changkyu , year =. Unveiling the Potential of Quantization with. 2603.08713 , archivePrefix =

    arXiv

  70. [70]

    Egiazarian, Vage and Castro, Roberto L. and Kuznedelev, Denis and Panferov, Andrei and Kurtic, Eldar and Pandit, Shubhra and Noll Marques, Alexandre and Kurtz, Mark and Ashkboos, Saleh and Hoefler, Torsten and Alistarh, Dan , booktitle =. Bridging the Gap Between Promise and Performance for Microscaling. 2026 , note =

    2026

  71. [71]

    2025 , pages =

    Lee, Janghwan and Park, Jiwoong and Kim, Jinseok and Kim, Yongjik and Oh, Jungju and Oh, Jinwook and Choi, Jungwook , booktitle =. 2025 , pages =. doi:10.18653/v1/2025.findings-acl.776 , url =

    doi:10.18653/v1/2025.findings-acl.776 2025

  72. [72]

    2026 , note =

    Liu, Wenyuan and Meng, Haoqian and Luo, Yilun and Zhang, Peng and Ma, Xindian , booktitle =. 2026 , note =

    2026

  73. [73]

    2026 , isbn =

    Hu, Weiming and Zhang, Zihan and Zhang, Haoyan and Zhang, Chen and Guo, Cong and Feng, Yu and Hu, Tianchi and Li, Guanglin and Hu, Guipeng and Wang, Junsong and Leng, Jingwen , title =. 2026 , isbn =. doi:10.1145/3779212.3790185 , booktitle =

    arXiv 2026

  74. [74]

    2601.07475 , archivePrefix =

    Meng, Haoqian and Luo, Yilun and Zhao, Yafei and Liu, Wenyuan and Zhang, Peng and Ma, Xindian , year =. 2601.07475 , archivePrefix =

    Pith/arXiv arXiv

  75. [75]

    Training

    Tseng, Albert and Yu, Tao and Park, Youngsuk , booktitle =. Training. 2025 , volume =

    2025

  76. [76]

    and Panferov, Andrei and Tabesh, Soroush and Sieberling, Oliver and Chen, Jiale and Nikdan, Mahdi and Ashkboos, Saleh and Alistarh, Dan , booktitle =

    Castro, Roberto L. and Panferov, Andrei and Tabesh, Soroush and Sieberling, Oliver and Chen, Jiale and Nikdan, Mahdi and Ashkboos, Saleh and Alistarh, Dan , booktitle =. Quartet: Native. 2025 , note =

    2025

  77. [77]

    2025 , note =

    Chmiel, Brian and Fishman, Maxim and Banner, Ron and Soudry, Daniel , booktitle =. 2025 , note =

    2025

  78. [78]

    2025 , eprint =

    Pretraining Large Language Models with. 2025 , eprint =

    2025

  79. [79]

    Oscillation-Reduced

    Chen, Yuxiang and Xi, Haocheng and Zhu, Jun and Chen, Jianfei , booktitle =. Oscillation-Reduced. 2025 , volume =

    2025

  80. [80]

    2209.05433 , archivePrefix =

    Micikevicius, Paulius and Stosic, Dusan and Burgess, Neil and Cornea, Marius and Dubey, Pradeep and Grisenthwaite, Richard and Ha, Sangwon and Heinecke, Alexander and Judd, Patrick and Kamalu, John and Mellempudi, Naveen and Oberman, Stuart and Shoeybi, Mohammad and Siu, Michael and Wu, Hao , year =. 2209.05433 , archivePrefix =

    Pith/arXiv arXiv

Showing first 80 references.