Pith. sign in

REVIEW 4 major objections 5 minor 66 references

First end-to-end 4-bit RL post-training: the bottleneck is rollout activations, not training precision.

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

2026-08-01 14:06 UTC pith:YGP6Y46H

load-bearing objection Novel first FP4 end-to-end RL result with a plausible rollout-activation diagnosis, but single-run numbers and simulated quantization mean the headline gains are not yet established. the 4 major comments →

arxiv 2607.26515 v1 pith:YGP6Y46H submitted 2026-07-29 cs.LG cs.AI

HiFloat4 Format for End-To-End Reinforcement Learning Post-Training of Large Language Models

Hei Yi Mak , Shadan Golestan , Hoang Le , Mehran Taghian Jazi , Yunke Peng , Yaoyuan Wang , Yao Wang , Junsong Wang
show 5 more authors
Tianchi Hu Fengchen He Guipeng Hu Tanzila Rahman Anandharaju Durai Raju
This is my paper
classification cs.LG cs.AI
keywords FP4 quantizationreinforcement learning post-trainingLLM reasoningrollout quantizationactivation outlierssparse residual correctionHiFloat4MXFP4
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.

The paper establishes that in fully FP4 reinforcement-learning post-training of LLMs, the dominant source of accuracy loss is not quantization of the training pass but quantization of rollout activations. Activation outliers stretch the dynamic range so that a large fraction of values underflow to zero under FP4, corrupting the rollout distribution. The authors show that keeping the training policy in higher precision while leaving the rollout in FP4 makes accuracy worse, ruling out the standard assumption that training-side noise is the problem. They introduce Rollout-ResQ, a sparse residual correction added only to the FP4 rollout matmul, which recovers most of the lost precision without adding dense compute. Combined with the HiFloat4 format's hierarchical scaling, the method closes the BF16 accuracy gap on GSM8K from 4.9% to 1.1% on Qwen2.5-3B, and works across both HiF4 and the open-standard MXFP4, indicating format choice sets the ceiling on recoverable accuracy.

Core claim

The paper's central claim is that end-to-end FP4 RL post-training is feasible if the rollout-side activation quantization error is corrected, and the principal failure mode is not training-side quantization but rollout-training mismatch caused by outlier-driven activation underflow. The evidence is a controlled precision matrix: with a BF16 rollout, FP4 training matches BF16 accuracy; with an FP4 rollout and BF16 training, accuracy collapses below the fully FP4 baseline. The mechanism is identified as 'zero collapse'—outliers force block scales so large that the majority of activation values round to zero. The proposed fix, Rollout-ResQ, computes a residual ΔX = X - Q(X), quantizes it, appli

What carries the argument

Rollout Residual Quantization (Rollout-ResQ) is the central mechanism: a single residual term attached to the FP4 rollout matrix multiplication, given by Ŷ = Q(X)Q(W)^T + S(Q(ΔX))Q(W)^T, where ΔX is the activation quantization residual and S is a hardware-friendly sparsity pattern (the paper uses S2:4). The sparse constraint is the key enabler: the residual is more robust to sparsification than the primary activation (measured via MSE/RMSE), so the correction runs at a fraction of a dense FP4 GEMM's cost. The companion object is the HiFloat4 (HiF4) format, whose three-level hierarchical scaling (E6M2 top scaler, E1 second/third scalers, S1P2/E1M2 elements over 64-element blocks) preserves re

Load-bearing premise

The paper's load-bearing premise is that its simulated FP4 quantization faithfully reproduces what real FP4 hardware will do; the authors explicitly state that no native FP4 hardware was available, so the accuracy gains and compute-efficiency estimates could fail to transfer if real kernels behave differently (e.g., different rounding, underflow handling, or sparse-kernel efficiency).

What would settle it

Run the same GRPO training with a real FP4 accelerator (or a cycle-accurate simulator that matches the target hardware's microarchitecture) and compare GSM8K-test accuracy and wall-clock time against the simulated results; if the accuracy gap to BF16 is not within ~1% or the speedup over FP16 falls below the predicted 2.67×, the central claims are falsified. Additionally, a larger number of random seeds (e.g., 5) would test whether the single-run ~4% effect is above run-to-run variance.

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

If this is right

  • Fully quantized FP4 RL post-training can reach within ~1% of BF16 accuracy on GSM8K with the proposed correction, suggesting 4-bit rollouts and training are viable for reasoning tasks.
  • The diagnosis reframes low-precision RL: the bottleneck is the rollout distribution's activation outliers, not the training pass's quantization noise, so future methods should focus on rollout-side corrections.
  • The sparse-residual design implies that adding a lightweight sparse GEMM to the rollout path can recover most of the precision lost to underflow while preserving theoretical 2.67× compute speedup over FP16.
  • Format choice is a major determinant of the accuracy ceiling: HiF4's hierarchical scaling outperforms MXFP4's single per-block scale by a wide margin (gap 4.9% vs 13.6% to BF16), so format design is as important as the correction method.
  • The method is format-agnostic: applying the same Rollout-ResQ recipe to MXFP4 narrows its gap from 13.6% to 5.3%, indicating the residual-correction principle transfers across FP4 formats.

Where Pith is reading between the lines

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

  • The zero-collapse mechanism implies a testable prediction: any FP4 format that reduces the fraction of blocks containing large outliers should reduce the required residual sparsity, possibly allowing denser or cheaper corrections.
  • If the diagnosis holds, then calibration-based outlier methods (SmoothQuant, RHT) are fundamentally mismatched to RL because activation statistics shift every iteration; a dynamic, per-tensor residual is the natural fit, suggesting future work on adaptive sparsity levels.
  • The authors' overhead analysis assumes a 4× FP4-vs-FP16 throughput advantage; on real hardware where the sparse GEMM kernel's efficiency differs, the actual speedup may be lower, so the claimed 2.67× is an upper bound pending native FP4 support.
  • The finding that BF16 training with FP4 rollout hurts more than full FP4 implies that mismatch itself destabilizes GRPO's importance ratios; this suggests that mixed-precision RL pipelines may be inherently fragile unless the rollout and training policies share the same numerical semantics.

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

4 major / 5 minor

Summary. The paper proposes, to its knowledge, the first end-to-end FP4 reinforcement-learning post-training pipeline for LLMs, with both rollout and training policies operating at 4-bit precision, and no LoRA or higher-precision gradient path. A component ablation (Table 1) is used to argue that the dominant degradation in FP4 RL comes from rollout-side activation quantization, specifically from outliers causing a large fraction of activations to underflow to zero under FP4. The paper then introduces Rollout-ResQ, a sparse residual correction applied only to the FP4 rollout matrix multiplications, and reports that with HiF4 it closes the BF16 gap on GSM8K-test from 4.9% to 1.1%, and that the same recipe applied to MXFP4 narrows the gap from 13.6% to 5.3%. Experiments are reported on Qwen2.5-3B with GSM8K and Qwen2.5-Math-7B with DAPO-Math-17B, using the VeRL framework and FP4 simulations. The paper also includes a theoretical overhead analysis claiming a 2.67x speedup over FP16 for the S2:4 variant, and concludes that HiF4 is the enabling format for fully FP4 RL post-training. Section 6 explicitly acknowledges that all experiments are FP4 simulations because native FP4 hardware is unavailable, and that runtime limits prevented tests on larger models.

Significance. If the central empirical claim holds, this would be a notable contribution to low-precision LLM RL: it identifies a specific failure mode (rollout activation zero collapse from outliers), proposes a simple and hardware-plausible correction, and provides a format-agnostic evaluation on both HiF4 and MXFP4. The paper's strengths include the internal consistency of the component ablation, the public code link, the explicit limitations section, and the fact that the method is validated on two different FP4 formats rather than being derived from a single format's parameters. However, the current evidence is not yet sufficient to fully establish the diagnosis: the key comparisons in Tables 1 and 2 are single-run, no seeds or confidence intervals are reported, and the efficiency claims are theoretical. If the main qualitative conclusions survive replication with seeds and error bars, the paper would be of significant practical interest for making RL post-training of LLMs more efficient.

major comments (4)
  1. [Tables 1 and 2, Section 5.2] The central diagnosis and headline result rest on single-run differences that are small relative to typical GRPO seed variance. In Table 1, the W4A16 vs W16A4 comparison is 85.22 vs 84.30 (a 0.92-point gap), which is used to claim that activation quantization dominates over weight quantization; the BF16-train/FP4-rollout vs FP4/FP4 comparison is 78.01 vs 82.03 (a 4.02-point gap), and Table 2's Rollout-ResQ-S2:4 HiF4 result is 85.90 vs BF16 86.96 (a 1.06-point gap). No seeds, standard deviations, or confidence intervals are reported anywhere, and Appendix A gives hyperparameters but no seed list. GRPO reward and final accuracy are known to vary by several points across seeds. If seed variance is even modest, the 0.92-point comparison could reverse and the 1.06-point closure could be within noise. Since the entire 'rollout activation quantization is the dominant failure mode' conclusion an
  2. [Section 5.4 and Section 6] The efficiency claim is theoretical rather than measured, and the cost model omits important components. Section 5.4 assumes FP4 provides a 4x throughput advantage over FP16 and computes an effective cost of 0.75mdn for Rollout-ResQ-S2:4, claiming a 2.67x speedup over FP16. However, the paper's own Section 6 states that all experiments rely on FP4 simulations and that simulation overhead dominates training time. The cost model also ignores the operations needed to form the residual: subtracting Q(X) from X, quantizing the residual, applying the sparsity function, and the extra memory traffic. Relative to the vanilla FP4 rollout (0.5mdn under the same assumptions), Rollout-ResQ-S2:4 adds a 50% effective FP4-GEMM cost, which is not obviously 'minimal overhead.' Additionally, the model uses m as the token count; in autoregressive rollout decoding, the per-step GEMM shape is fundamentally di
  3. [Abstract, Tables 2 and 3] The headline claim that Rollout-ResQ 'closes the accuracy gap to BF16 from 4.9% to 1.1%' is reported in the abstract without stating that this applies only to Qwen2.5-3B on GSM8K-test. The same paper's Table 2 shows a larger remaining gap on Math-500 (8.6 points for HiF4 under Rollout-ResQ-S2:4), and Table 3 for Qwen2.5-Math-7B shows substantial remaining gaps across all four benchmarks even with the best variant: AIME24 6.56, AIME25 3.54, AMC23 8.75, and Math500 5.64 points. The general statement in the abstract and conclusion that the method brings fully quantized FP4 RL 'within striking distance of full precision' is therefore supported only by one selected benchmark. Please qualify the headline claim to the specific benchmark and model, and discuss the benchmark-dependent variation in the remaining gap.
  4. [Table 3, Section 5.3] The paper generalizes the 3B diagnosis to the 7B model without providing the corresponding component ablation for 7B. Table 1, which isolates the effects of training quantization, rollout quantization, and activation-only versus weight-only quantization, is reported only for Qwen2.5-3B on GSM8K. Table 3 for Qwen2.5-Math-7B reports only the full BF16 baseline, the full HiF4 baseline, and recovery methods; there is no 7B row with BF16 training/FP4 rollout or with W4A16/W16A4 to confirm that rollout activation quantization is still the dominant failure mode at 7B scale, where training dynamics and outlier behavior can differ. The conclusion that the rollout-side correction is the right remedy for 7B therefore depends on an extrapolation from a much smaller model trained on a simpler task. Adding at least one 7B comparison of training-side vs rollout-side quantization would substantially str
minor comments (5)
  1. [Section 1] Typo: 'performsworse' should be 'performs worse.' Also, 'a systematic study reveals' and similar phrases in the abstract are repeated verbatim in the introduction; consider tightening.
  2. [Section 5.3 / Figure 5 caption] The dataset name is inconsistent: Section 5.1 and the text refer to 'DAPO-Math-17B' for Qwen2.5-Math-7B, while Section 5.3's title and Figure 5(c)'s caption use 'DAPO-Math-17K' and 'DAPO-Math-17K' appears in Figure 5. Please unify the notation.
  3. [Equation (2)] The GRPO objective is stated without the KL penalty, and the text says the KL term is omitted; this is acceptable, but the sentence 'Note that, the objective also contains...' has a stray comma and should read 'Note that the objective also contains...'.
  4. [References] Reference [10] has a formatting issue: 'Gpt3. int8 ()' should be 'GPT-3. int8().' Reference [16] has a name in all caps ('Daniel HAZIZA'); please fix to title case.
  5. [Section 5.2, Figure 5] The text says 'The figure reports the best-performing variant of Rollout-ResQ, i.e., Rollout-ResQ-S2:4,' but Tables 2 and 3 show that S2:4 is not strictly the best on all benchmarks (e.g., Math-500 under HiF4). Please clarify that S2:4 is selected as the overall recommended variant, not the best on every metric.

Circularity Check

0 steps flagged

No significant circularity: controlled Table 1/2 comparisons and the Rollout-ResQ construction are independent of their inputs; HiF4 self-citations are background, not load-bearing.

full rationale

The paper's central diagnosis—rollout activation quantization, not training quantization, dominates FP4 RL degradation—is derived from Table 1's precision-ablation rows (FP4/BF16, W4A16 vs W16A4, BF16-train/FP4-rollout vs FP4/FP4). These are empirical comparisons, not quantities that equal each other by construction. Rollout-ResQ is explicitly defined as a residual correction ΔXℓ = Xℓ − Q(Xℓ) (Eq. 3) and its projection (Eq. 5) is a sum of a quantized matmul and a sparsified residual matmul; this is a constructive approximation to XℓWℓ^T, and its sparsity choices (α=50, S2:4) are motivated by prior hardware/pruning literature, not fitted to the target accuracies. The paper builds on HiF4 [36] and cites HiF4 pretraining [49], both with overlapping authors, but the recommendation of HiF4 is supported by in-paper head-to-head comparisons against MXFP4 in Tables 2 and 3; removing those self-citations would not collapse the argument. The acknowledged limitation that all experiments rely on FP4 simulation, and the absence of seeds/confidence intervals, are external-validity and statistical concerns, not circularity. I therefore find no step where a predicted result reduces by definition to a fitted input or to a self-citation chain.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

The central accuracy result depends mainly on hand-chosen sparsity levels (50%) and on the fidelity of FP4 simulation. No new physical or mathematical entities are introduced; Rollout-ResQ is an algorithmic correction. The paper's own limitation statement (Section 6) confirms the simulation dependence.

free parameters (2)
  • Rollout-ResQ sparsity level = 50% (S2:4, S50%, S32x64)
    Chosen by hand in Sec 4.3 ('we choose α=50', 'We choose S2:4'); no sensitivity sweep is reported, and the headline results depend on this level.
  • Variant selection for headline = S2:4
    S2:4 is reported as best on GSM8K-test, but other variants win on other benchmarks (e.g., S32x64 on Math500 under HiF4, Dense on some 7B tasks). The single-number headline is a post-hoc selection across variants.
axioms (3)
  • domain assumption FP4 simulation faithfully reproduces future hardware numerics.
    Section 6 admits there is no native FP4 hardware and all experiments rely on FP4 simulations; if the simulator differs from real kernels, central accuracy and efficiency claims may not transfer.
  • domain assumption Block-scale outlier dominance is the mechanism for zero collapse.
    Figures 1 and 2 show correlation between outlier blocks and post-quantization zeros, but no formal proof that outlier-driven scale expansion is the sole or main cause across all layers.
  • domain assumption FP4 GEMM provides 4x throughput over FP16 on supported hardware.
    Section 5.4 assumes a 4x advantage based on cited hardware references; no measurement is made in this paper because runs are simulated.

pith-pipeline@v1.3.0-daily-deepseek · 17366 in / 14058 out tokens · 112843 ms · 2026-08-01T14:06:43.697395+00:00 · methodology

0 comments
read the original abstract

We present, to our knowledge, the first end-to-end FP4 RL post-training, in which both the rollout and training policies, including their forward and backward passes, operate at 4-bit precision. A systematic study reveals that the dominant source of degradation in FP4 RL is not training-side quantization error but rollout activation quantization: outliers stretch the dynamic range so far that a large number of activation values underflow to zero under FP4. Counterintuitively, restoring the training policy to higher precision while keeping the rollout in FP4 makes accuracy worse than full FP4 baseline, exposing rollout-training mismatch as the principal failure mode and ruling out standard pretraining-style fixes. We address this with Rollout Residual Quantization (Rollout-ResQ): a single residual correction term constrained to a hardware-friendly sparsity pattern, added only to the FP4 rollout matmul -- a lightweight correction that recovers most of the precision lost to outlier-driven underflow without inflating the rollout's compute footprint. On Qwen2.5-3B and Qwen2.5-Math-7B, Rollout-ResQ paired with the HiFloat4 (HiF4) format -- whose three-level hierarchical scaling preserves resolution under FP4's tight 4-bit budget -- closes the accuracy gap to BF16 from 4.9% to 1.1%, bringing fully quantized FP4 RL within striking distance of full precision. Applied to the open-standard MXFP4, the same recipe narrows the gap from 13.6% to 5.3%, revealing that FP4 format choice is a key factor that determines the ceiling on recoverable accuracy. Together, these results establish HiF4 as the enabling format for end-to-end FP4 RL post-training, and Rollout-ResQ as the activation-side mechanism that makes the gap to BF16 closable.

Figures

Figures reproduced from arXiv: 2607.26515 by Anandharaju Durai Raju, Fengchen He, Guipeng Hu, Hei Yi Mak, Hoang Le, Junsong Wang, Mehran Taghian Jazi, Shadan Golestan, Tanzila Rahman, Tianchi Hu, Yao Wang, Yaoyuan Wang, Yunke Peng.

Figure 1
Figure 1. Figure 1: Percentage of zero values in activa￾tion tensors of each layer before quantization (BF16) and after quantization (HiF4, MXFP4) with GSM8K and Qwen2.5-3B. 0 4 8 12 16 20 24 28 32 0 20 40 60 80 100 % blocks with outliers down_projection block=64 block=32 0 4 8 12 16 20 24 28 32 gate_up_projection block=64 block=32 0 4 8 12 16 20 24 28 32 Transformer block index 0 20 40 60 80 100 % blocks with outliers o_proj… view at source ↗
Figure 3
Figure 3. Figure 3: Rollout-ResQ overview. dramatic increase in sparsity. In specific layers, such as the down projection of transformer block 1 and 2, sparsity exceeds 80%, indicating a "zero collapse" where the majority of non-zero values are mapped to zero. This phenomenon is driven by the presence of outliers, as emphasized in [51] and [64]. As shown in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Heatmap of quantized activation and residual during rollout. The MSE and RMSE be￾fore and after performing 2:4 sparsity. The choice to sparsify the residual rather than the primary activation is motivated by their ro￾bustness to sparsification. We measure MSE and RMSE of Q(Xℓ) and Q(∆Xℓ) with their corre￾sponding S2:4 sparsified versions. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Training performance (average reward) for Qwen2.5-3B on GSM8K under HiF4 and [PITH_FULL_IMAGE:figures/full_fig_p008_5.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

66 extracted references · 14 linked inside Pith

  1. [1]

    Pareto-optimal quantized resnet is mostly 4-bit

    Shivani Agrawal Jonathan Malmaud Oleg Rybakov Chas Leichner AmirAli Abdolrashidi, Lisa Wang and Lukasz Lew. Pareto-optimal quantized resnet is mostly 4-bit. In2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2021. URL http://dx.doi.org/10.1109/CVPRW53098.2021.00345

    arXiv 2021

  2. [2]

    Structured pruning of deep convolutional neural networks.J

    Sajid Anwar, Kyuyeon Hwang, and Wonyong Sung. Structured pruning of deep convolutional neural networks.J. Emerg. Technol. Comput. Syst., 13(3), February 2017. ISSN 1550-4832. doi: 10.1145/3005348. URLhttps://doi.org/10.1145/3005348

    doi:10.1145/3005348 2017

  3. [3]

    Quarot: Outlier-free 4-bit inference in rotated llms.Advances in Neural Information Processing Systems, 37:100213– 100240, 2024

    Saleh Ashkboos, Amirkeivan Mohtashami, Maximilian L Croci, Bo Li, Pashmina Cameron, Martin Jaggi, Dan Alistarh, Torsten Hoefler, and James Hensman. Quarot: Outlier-free 4-bit inference in rotated llms.Advances in Neural Information Processing Systems, 37:100213– 100240, 2024

    2024

  4. [4]

    Reflective multi-agent collaboration based on large language models

    Xiaohe Bo, Zeyu Zhang, Quanyu Dai, Xueyang Feng, Lei Wang, Rui Li, Xu Chen, and Ji- Rong Wen. Reflective multi-agent collaboration based on large language models. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=wWiAR5mqXq

    2024

  5. [5]

    Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, et al

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, et al. Language models are few-shot learners.Advances in Neural Information Processing Systems, 2020

    2020

  6. [6]

    PDP: Parameter-free differentiable pruning is all you need

    Minsik Cho, Saurabh Adya, and Devang Naik. PDP: Parameter-free differentiable pruning is all you need. InThirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/forum?id=lLztVBaBVU

    2023

  7. [7]

    Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021

    Pith/arXiv arXiv 2021

  8. [8]

    Binarized neural networks: Training deep neural networks with weights and activations constrained to+ 1 or-1.arXiv preprint arXiv:1602.02830, 2016

    Matthieu Courbariaux, Itay Hubara, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized neural networks: Training deep neural networks with weights and activations constrained to+ 1 or-1.arXiv preprint arXiv:1602.02830, 2016

    Pith/arXiv arXiv 2016

  9. [9]

    Microscaling data formats for deep learning

    Bita Darvish Rouhani, Ritchie Zhao, Ankit More, Mathew Hall, Alireza Khodamoradi, Summer Deng, Dhruv Choudhary, Marius Cornea, Eric Dellinger, Kristof Denolf, Stosic Dusan, Venmugil Elango, Maximilian Golub, Alexander Heinecke, Phil James-Roxby, Dharmesh Jani, Gaurav Kolhe, Martin Langhammer, Ada Li, Levi Melnick, Maral Mes- makhosroshahi, Andres Rodrigue...

    2023

  10. [10]

    Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. Gpt3. int8 (): 8-bit matrix multiplication for transformers at scale.Advances in neural information processing systems, 35: 30318–30332, 2022

    2022

  11. [11]

    Svirschevski, Vage Egiazarian, Denis Kuznedelev, Elias Frantar, Saleh Ashkboos, Alexander Borzunov, Torsten Hoefler, and Dan Alistarh

    Tim Dettmers, Ruslan A. Svirschevski, Vage Egiazarian, Denis Kuznedelev, Elias Frantar, Saleh Ashkboos, Alexander Borzunov, Torsten Hoefler, and Dan Alistarh. SpQR: A sparse-quantized representation for near-lossless LLM weight compression. InThe Twelfth International Con- ference on Learning Representations, 2024. URL https://openreview.net/forum?id= Q1u25ahSuy

    2024

  12. [12]

    From problem-solving to teaching problem-solving: Aligning llms with pedagogy using reinforcement learning

    David Dinucu-Jianu, Jakub Macina, Nico Daheim, Ido Hakimi, Iryna Gurevych, and Mrinmaya Sachan. From problem-solving to teaching problem-solving: Aligning llms with pedagogy using reinforcement learning. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 272–292, 2025. 10

    2025

  13. [13]

    Chao Fang, Aojun Zhou, and Zhongfeng Wang. An Algorithm–Hardware Co-Optimized Framework for Accelerating N:M Sparse Transformers .IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 30(11):1573–1586, November 2022. ISSN 1557-9999. doi: 10.1109/TVLSI.2022.3197282. URL https://doi.ieeecomputersociety.org/10. 1109/TVLSI.2022.3197282

    arXiv 2022

  14. [14]

    Megablocks: Ef- ficient sparse training with mixture-of-experts

    Trevor Gale, Deepak Narayanan, Cliff Young, and Matei Zaharia. Megablocks: Ef- ficient sparse training with mixture-of-experts. In D. Song, M. Carbin, and T. Chen, editors,Proceedings of Machine Learning and Systems, volume 5, pages 288–304. Cu- ran, 2023. URL https://proceedings.mlsys.org/paper_files/paper/2023/file/ 5a54f79333768effe7e8927bcccffe40-Pape...

    2023

  15. [15]

    Would i lie to you? inference time alignment of language models using direct preference heads

    Avelina Asada Hadji-Kyriacou and Ognjen Arandjelovic. Would i lie to you? inference time alignment of language models using direct preference heads. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview. net/forum?id=NKGuLthW80

    2024

  16. [16]

    Accelerating transformer inference and training with 2:4 activation sparsity

    Daniel HAZIZA, Timothy Chou, Dhruv Choudhary, Jesse Cai, Luca Wehrstedt, Francisco Massa, Jiecao Yu, Geonhwa Jeong, Supriya Rao, and Patrick Labatut. Accelerating transformer inference and training with 2:4 activation sparsity. InSparsity in LLMs (SLLM): Deep Dive into Mixture of Experts, Quantization, Hardware, and Inference, 2025. URLhttps://openreview....

    2025

  17. [17]

    Measuring mathematical problem solving with the MATH dataset

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the MATH dataset. InThirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. URLhttps://openreview.net/forum?id=7Bywt2mQsCe

    2021

  18. [18]

    OpenRLHF: A ray-based easy-to-use, scalable and high-performance RLHF frame- work

    Jian Hu, Xibin Wu, Wei Shen, Jason Klein Liu, Weixun Wang, Songlin Jiang, Haoran Wang, Hao Chen, Bin Chen, Wenkai Fang, Xianyu, Yu Cao, Haotian Xu, and Yiming Liu. OpenRLHF: A ray-based easy-to-use, scalable and high-performance RLHF frame- work. In Ivan Habernal, Peter Schulam, and Jörg Tiedemann, editors,Proceedings of the 2025 Conference on Empirical M...

    2025

  19. [19]

    Accelerating transformer pre-training with 2:4 sparsity

    Yuezhou Hu, Kang Zhao, Weiyu Huang, Jianfei Chen, and Jun Zhu. Accelerating transformer pre-training with 2:4 sparsity. InProceedings of the 41st International Conference on Machine Learning, ICML’24. JMLR.org, 2024

    2024

  20. [20]

    QeRL: Beyond efficiency - quantization-enhanced reinforcement learning for LLMs

    Wei Huang, Yi Ge, Shuai Yang, Yicheng Xiao, Huizi Mao, Yujun Lin, Hanrong Ye, Sifei Liu, Ka Chun Cheung, Hongxu Yin, Yao Lu, Xiaojuan Qi, Song Han, and Yukang Chen. QeRL: Beyond efficiency - quantization-enhanced reinforcement learning for LLMs. InThe Fourteenth International Conference on Learning Representations, 2026. URL https://openreview. net/forum?...

    2026

  21. [21]

    Aime_2024

    Maxwell Jia. Aime_2024. Hugging Face dataset, 2024. https://huggingface.co/ datasets/Maxwell-Jia/AIME_2024

    2024

  22. [22]

    Pact: Parameterized clipping activation for quantized neural networks.arXiv preprint arXiv:1805.06085, 2018

    Swagath Venkataramani Pierce I-Jen Chuang Vijayalakshmi Srinivasan Kailash Gopalakrishnan Jungwook Choi, Zhuo Wang. Pact: Parameterized clipping activation for quantized neural networks.arXiv preprint arXiv:1805.06085, 2018

    Pith/arXiv arXiv 2018

  23. [23]

    Mahoney, and Kurt Keutzer

    Sehoon Kim, Coleman Richard Charles Hooper, Amir Gholami, Zhen Dong, Xiuyu Li, Sheng Shen, Michael W. Mahoney, and Kurt Keutzer. SqueezeLLM: Dense-and-sparse quantization. In Forty-first International Conference on Machine Learning, 2024. URL https://openreview. net/forum?id=0jpbpFia8m

    2024

  24. [24]

    A systematic survey and critical review on evaluating large language models: Challenges, limitations, and recommendations

    Md Tahmid Rahman Laskar, Sawsan Alqahtani, M Saiful Bari, Mizanur Rahman, Mohammad Abdullah Matin Khan, Haidar Khan, Israt Jahan, Amran Bhuiyan, Chee Wei Tan, Md Rizwan Parvez, et al. A systematic survey and critical review on evaluating large language models: Challenges, limitations, and recommendations. InProceedings of the 2024 Conference on Empirical ...

    2024

  25. [25]

    Owq: Outlier- aware weight quantization for efficient fine-tuning and inference of large language models

    Changhun Lee, Jungyu Jin, Taesu Kim, Hyungjun Kim, and Eunhyeok Park. Owq: Outlier- aware weight quantization for efficient fine-tuning and inference of large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 13355–13364, 2024

    2024

  26. [26]

    QuRL: Low-precision reinforcement learning for efficient reasoning

    Yuhang Li, Reena Elangovan, Xin Dong, Priyadarshini Panda, and Brucek Khailany. QuRL: Low-precision reinforcement learning for efficient reasoning. InThe Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum? id=eG0bpCwdKn

    2026

  27. [27]

    Enhancing efficiency and exploration in reinforcement learning for llms

    Mengqi Liao, Xiangyu Xi, Chen Ruinian, Jia Leng, Yangen Hu, Ke Zeng, Shuai Liu, and Huaiyu Wan. Enhancing efficiency and exploration in reinforcement learning for llms. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 1451–1463, 2025

    2025

  28. [28]

    Duquant: Distributing outliers via dual transformation makes stronger quantized LLMs

    Haokun Lin, Haobo Xu, Yichen Wu, Jingzhi Cui, Yingtao Zhang, Linzhan Mou, Linqi Song, Zhenan Sun, and Ying Wei. Duquant: Distributing outliers via dual transformation makes stronger quantized LLMs. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URLhttps://openreview.net/forum?id=mp8u2Pcmqz

    2024

  29. [29]

    Awq: Activation-aware weight quantization for on-device llm compression and acceleration.Proceedings of machine learning and systems, 6:87–100, 2024

    Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. Awq: Activation-aware weight quantization for on-device llm compression and acceleration.Proceedings of machine learning and systems, 6:87–100, 2024

    2024

  30. [30]

    Flashrl: 8bit rollouts, full power rl, August 2025

    Liyuan Liu, Feng Yao, Dinghuai Zhang, Chengyu Dong, Jingbo Shang, and Jianfeng Gao. Flashrl: 8bit rollouts, full power rl, August 2025. URL https://fengyao.notion.site/ flash-rl. Preprint

    2025

  31. [31]

    Gdpo: Group reward- decoupled normalization policy optimization for multi-reward rl optimization.arXiv preprint arXiv:2601.05242, 2026

    Shih-Yang Liu, Xin Dong, Ximing Lu, Shizhe Diao, Peter Belcak, Mingjie Liu, Min-Hung Chen, Hongxu Yin, Yu-Chiang Frank Wang, Kwang-Ting Cheng, et al. Gdpo: Group reward- decoupled normalization policy optimization for multi-reward rl optimization.arXiv preprint arXiv:2601.05242, 2026

    Pith/arXiv arXiv 2026

  32. [32]

    Nover: Incentive training for language models via verifier-free reinforcement learning

    Wei Liu, Siya Qi, Xinyu Wang, Chen Qian, Yali Du, and Yulan He. Nover: Incentive training for language models via verifier-free reinforcement learning. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 7450–7469, 2025

    2025

  33. [33]

    Spinquant: LLM quantization with learned rotations

    Zechun Liu, Changsheng Zhao, Igor Fedorov, Bilge Soran, Dhruv Choudhary, Raghuraman Krishnamoorthi, Vikas Chandra, Yuandong Tian, and Tijmen Blankevoort. Spinquant: LLM quantization with learned rotations. InThe Thirteenth International Conference on Learning Representations, 2025. URLhttps://openreview.net/forum?id=ogO6DGE6FZ

    2025

  34. [34]

    Whatmough, Yuhao Zhu, and Matthew Mattina

    Zhi Gang Liu, Paul N. Whatmough, Yuhao Zhu, and Matthew Mattina. S2TA: exploiting structured sparsity for energy-efficient mobile CNN acceleration. InIEEE International Symposium on High-Performance Computer Architecture, HPCA 2022, Seoul, South Korea, April 2-6, 2022, pages 573–586. IEEE, 2022. doi: 10.1109/HPCA53966.2022.00049. URL https://doi.org/10.11...

    arXiv 2022

  35. [35]

    Understanding r1-zero-like training: A critical perspective

    Zichen Liu, Changyu Chen, Wenjun Li, Penghui Qi, Tianyu Pang, Chao Du, Wee Sun Lee, and Min Lin. Understanding r1-zero-like training: A critical perspective. InSecond Conference on Language Modeling, 2025. URLhttps://openreview.net/forum?id=5PAF7PAY2Y

    2025

  36. [36]

    Hifloat4 format for language model inference.arXiv preprint arXiv:2602.11287, 2026

    Yuanyong Luo, Jing Huang, Yu Cheng, Ziwei Yu, Kaihua Zhang, Kehong Hong, Xinda Ma, Xin Wang, Anping Tong, Guipeng Hu, et al. Hifloat4 format for language model inference.arXiv preprint arXiv:2602.11287, 2026

    arXiv 2026

  37. [37]

    Hrbp: Hardware-friendly regrouping towards block-based pruning for sparse cnn training

    Haoyu Ma, Chengming Zhang, lizhi xiang, Xiaolong Ma, Geng Yuan, Wenkai Zhang, Shiwei Liu, Tianlong Chen, Dingwen Tao, Yanzhi Wang, Zhangyang Wang, and Xiao- hui Xie. Hrbp: Hardware-friendly regrouping towards block-based pruning for sparse cnn training. In Yuejie Chi, Gintare Karolina Dziugaite, Qing Qu, Atlas Wang Wang, and Zhihui Zhu, editors,Conference...

    2024

  38. [38]

    Aime 2025

    MathArena. Aime 2025. Hugging Face dataset, 2025. https://huggingface.co/datasets/MathArena/aime_2025

    2025

  39. [39]

    Arc- quant: Boosting nvfp4 quantization with augmented residual channels for llms.arXiv preprint arXiv:2601.07475, 2026

    Haoqian Meng, Yilun Luo, Yafei Zhao, Wenyuan Liu, Peng Zhang, and Xindian Ma. Arc- quant: Boosting nvfp4 quantization with augmented residual channels for llms.arXiv preprint arXiv:2601.07475, 2026

    Pith/arXiv arXiv 2026

  40. [40]

    Accelerating sparse deep neural networks, 2021

    Asit Mishra, Jorge Albericio Latorre, Jeff Pool, Darko Stosic, Dusan Stosic, Ganesh Venkatesh, Chong Yu, and Paulius Micikevicius. Accelerating sparse deep neural networks, 2021. URL https://arxiv.org/abs/2104.08378

    Pith/arXiv arXiv 2021

  41. [41]

    Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feed- bac...

    2022

  42. [42]

    Channel permutations for n:m sparsity

    Jeff Pool and Chong Yu. Channel permutations for n:m sparsity. In M. Ranzato, A. Beygelzimer, Y . Dauphin, P.S. Liang, and J. Wortman Vaughan, editors,Advances in Neural Information Processing Systems, volume 34, pages 13316–13327. Curran Associates, Inc., 2021. URL https://proceedings.neurips.cc/paper_files/paper/2021/file/ 6e8404c3b93a9527c8db241a184659...

    2021

  43. [43]

    Qwen2.5 technical report, 2025

    Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li,...

    Pith/arXiv arXiv 2025

  44. [44]

    Ocp microscaling (mx) specification

    Bita Darvish Rouhani, Nitin Garegrat, Tom Savell, Ankit More, Kyung-Nam Han, Mathew Zhao, Ritchie Hall, Jasmine Klar, Eric Chung, Yuan Yu, Michael Schulte, Ralph Wittig, Ian Bratt, Nigel Stephens, Jelena Milanovic, John Brothers, Pradeep Dubey, Marius Cornea, Alexander Heinecke, Andres Rodriguez, Martin Langhammer, Summer Deng, Maxim Naumov, Paulius Micik...

    2023

  45. [45]

    Omniquant: Omnidirectionally calibrated quan- tization for large language models

    Wenqi Shao, Mengzhao Chen, Zhaoyang Zhang, Peng Xu, Lirui Zhao, Zhiqian Li, Kaipeng Zhang, Peng Gao, Yu Qiao, and Ping Luo. Omniquant: Omnidirectionally calibrated quan- tization for large language models. InThe Twelfth International Conference on Learning Representations, 2024. URLhttps://openreview.net/forum?id=8Wuvhh0LYW

    2024

  46. [46]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    Pith/arXiv arXiv 2024

  47. [47]

    Hybridflow: A flexible and efficient rlhf framework.arXiv preprint arXiv: 2409.19256, 2024

    Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. Hybridflow: A flexible and efficient rlhf framework.arXiv preprint arXiv: 2409.19256, 2024

    Pith/arXiv arXiv 2024

  48. [48]

    Fastcurl: Curriculum reinforcement learning with stage-wise context scaling for efficient training r1-like reasoning models.arXiv preprint arXiv:2503.17287, 2025

    Mingyang Song, Mao Zheng, Zheng Li, Wenjie Yang, Xuan Luo, Yue Pan, and Feng Zhang. Fastcurl: Curriculum reinforcement learning with stage-wise context scaling for efficient training r1-like reasoning models.arXiv preprint arXiv:2503.17287, 2025

    arXiv 2025

  49. [49]

    Hifloat4 format for language model pre-training on ascend npus, 2026

    Mehran Taghian, Yunke Peng, Xing Huang, Yao Wang, Yaoyuan Wang, Wei Guo, Yuanyong Luo, Tianchi Hu, Junsong Wang, Xin Wang, Hu Liu, Yu Cheng, Ziwei Yu, Hongliang Li, Mehdi Rahimifar, Lei Yan, Xuefei Wang, Zhuang Ma, Lei Liu, Hui Yu, Anandharaju Durai Raju, Hoang Le, Hei Yi Mak, Tanzila Rahman, and Shadan Golestan. Hifloat4 format for language model pre-tra...

    Pith/arXiv arXiv 2026

  50. [50]

    Training llms with mxfp4, 2025

    Albert Tseng, Tao Yu, and Youngsuk Park. Training llms with mxfp4, 2025. URL https: //arxiv.org/abs/2502.20586

    Pith/arXiv arXiv 2025

  51. [51]

    Optimizing large language model training using FP4 quan- tization

    Ruizhe Wang, Yeyun Gong, Xiao Liu, Guoshuai Zhao, Ziyue Yang, Baining Guo, Zheng- Jun Zha, and Peng CHENG. Optimizing large language model training using FP4 quan- tization. InForty-second International Conference on Machine Learning, 2025. URL https://openreview.net/forum?id=uK7JArZEJM

    2025

  52. [52]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems, volume 35, pages 24824–24837. Curran Assoc...

    2022

  53. [53]

    Webagent-r1: Training web agents via end-to-end multi-turn reinforcement learning

    Zhepei Wei, Wenlin Yao, Yao Liu, Weizhi Zhang, Qin Lu, Liang Qiu, Changlong Yu, Puyang Xu, Chao Zhang, Bing Yin, et al. Webagent-r1: Training web agents via end-to-end multi-turn reinforcement learning. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 7920–7939, 2025

    2025

  54. [54]

    Jet-rl: Enabling on-policy fp8 reinforcement learning with unified training and rollout precision flow.arXiv preprint arXiv:2601.14243, 2026

    Haocheng Xi, Charlie Ruan, Peiyuan Liao, Yujun Lin, Han Cai, Yilong Zhao, Shuo Yang, Kurt Keutzer, Song Han, and Ligeng Zhu. Jet-rl: Enabling on-policy fp8 reinforcement learning with unified training and rollout precision flow.arXiv preprint arXiv:2601.14243, 2026

    arXiv 2026

  55. [55]

    Smoothquant: Accurate and efficient post-training quantization for large language models

    Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han. Smoothquant: Accurate and efficient post-training quantization for large language models. InInternational conference on machine learning, pages 38087–38099. PMLR, 2023

    2023

  56. [56]

    Ultra-low precision 4-bit training of deep neural networks

    Chia-Yu Chen Jiamin Ni Ankur Agrawal Xiaodong Cui Swagath Venkataramani Kaoutar El Maghraoui Vijayalakshmi (Viji) Srinivasan Xiao Sun, Naigang Wang and Kailash Gopalakr- ishnan. Ultra-low precision 4-bit training of deep neural networks. InThirty-Fourth Annual Conference on Neural Information Processing Systems, 2020. URL https://openreview. net/forum?id=...

    2020

  57. [57]

    Logic-rl: Unleashing llm reasoning with rule-based reinforcement learning.arXiv preprint arXiv:2502.14768, 2025

    Tian Xie, Zitian Gao, Qingnan Ren, Haoming Luo, Yuqian Hong, Bryan Dai, Joey Zhou, Kai Qiu, Zhirong Wu, and Chong Luo. Logic-rl: Unleashing llm reasoning with rule-based reinforcement learning.arXiv preprint arXiv:2502.14768, 2025

    Pith/arXiv arXiv 2025

  58. [58]

    Qwen2.5-math technical report: Toward mathematical expert model via self-improvement, 2024

    An Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, Chengpeng Li, Dayiheng Liu, Jianhong Tu, Jingren Zhou, Junyang Lin, Keming Lu, Mingfeng Xue, Runji Lin, Tianyu Liu, Xingzhang Ren, and Zhenru Zhang. Qwen2.5-math technical report: Toward mathematical expert model via self-improvement, 2024. URLhttps://arxiv.org/abs/2409.12122

    Pith/arXiv arXiv 2024

  59. [59]

    On the rollout-training mismatch in modern RL systems

    Feng Yao, Liyuan Liu, Dinghuai Zhang, Chengyu Dong, Jingbo Shang, and Jianfeng Gao. On the rollout-training mismatch in modern RL systems. InOPT 2025: Optimization for Machine Learning, 2025. URLhttps://openreview.net/forum?id=xsonlP8DRV

    2025

  60. [60]

    DAPO: An open-source LLM reinforcement learning system at scale

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, YuYue, Weinan Dai, Tiantian Fan, Gaohong Liu, Juncai Liu, LingJun Liu, Xin Liu, Haibin Lin, Zhiqi Lin, Bole Ma, Guangming Sheng, Yuxuan Tong, Chi Zhang, Mofan Zhang, Ru Zhang, Wang Zhang, Hang Zhu, Jinhua Zhu, Jiaze Chen, Jiangjie Chen, Chengyi Wang, Hongli Yu, Yuxuan Song, Xiangpeng Wei, Hao ...

    2026

  61. [61]

    Dpquant: Efficient and differentially-private model training via dynamic quantization scheduling

    Gennady Pekhimenko Nandita Vijaykumar Yubo Gao, Renbo Tu. Dpquant: Efficient and differentially-private model training via dynamic quantization scheduling. InThe Fourteenth International Conference on Learning Representations, 2026. URL https://openreview. net/forum?id=neaxYXGYd5

    2026

  62. [62]

    Grpo-lead: A difficulty-aware reinforcement learning approach for concise mathematical reasoning in language models

    Jixiao Zhang and Chunsheng Zuo. Grpo-lead: A difficulty-aware reinforcement learning approach for concise mathematical reasoning in language models. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 5642–5665, 2025. 14

    2025

  63. [63]

    Rearank: Reasoning re- ranking agent via reinforcement learning

    Le Zhang, Bo Wang, Xipeng Qiu, Siva Reddy, and Aishwarya Agrawal. Rearank: Reasoning re- ranking agent via reinforcement learning. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 2458–2471, 2025

    2025

  64. [64]

    Accurate int8 training through dynamic block-level fallback.arXiv preprint arXiv:2503.08040, 2025

    Pengle Zhang, Jia Wei, Jintao Zhang, Jun Zhu, and Jianfei Chen. Accurate int8 training through dynamic block-level fallback.arXiv preprint arXiv:2503.08040, 2025

    Pith/arXiv arXiv 2025

  65. [65]

    Qspec: Speculative decoding with complementary quantization schemes

    Juntao Zhao, Wenhao Lu, Sheng Wang, Lingpeng Kong, and Chuan Wu. Qspec: Speculative decoding with complementary quantization schemes. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 4779–4795, 2025

    2025

  66. [66]

    Chi, Denny Zhou, Swaroop Mishra, and Steven Zheng

    Pei Zhou, Jay Pujara, Xiang Ren, Xinyun Chen, Heng-Tze Cheng, Quoc V Le, Ed H. Chi, Denny Zhou, Swaroop Mishra, and Steven Zheng. SELF-DISCOVER: Large language models self- compose reasoning structures. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URLhttps://openreview.net/forum?id=BROvXhmzYK. A Training Details GR...

    2024