Pith. sign in

REVIEW 4 major objections 5 minor 33 references

Recurrent Residual Quantization: A Progressive Multi-Precision Representation for LLMs

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read One quantized checkpoint yields 2-, 4-, 6-, and 8-bit versions of an LLM.

desk verdict Useful multi-precision PTQ wrapper with honest evaluation and credible 6/8-bit prefixes, but the outlier theory is an unverified post hoc story and the 4-bit results show real limits. read the letter →

arxiv 2608.04048 v1 pith:JWYASC6G submitted 2026-08-04 cs.LG cs.AI

Yu Luo , Bo Dong , Wenhua Cheng , Haihao Shen This is my paper · ORCID
classification cs.LGcs.AI
keywords recurrentresidualquantizationmulti-precisionpost-trainingLLMweightcompressionsingle-checkpointround-to-nearestoutlier-aware
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Recurrent Residual Quantization (RRQ) is a post-training scheme that stores an LLM's weights as a 2-bit quantized base tensor plus three 2-bit quantized residual tensors. Any prefix of the four stages is a usable model, so one checkpoint yields 2-, 4-, 6-, and 8-bit operating points. The claimed payoff is flexibility: a single stored package can be served at whatever precision a deployment's memory and latency allow, without building or storing separate checkpoints. The paper reports that the all-RTN version, built with no calibration data and no Hessian computation, is competitive with a joint multi-bit optimization baseline at 6 and 8 bits across six LLMs and is roughly 3.3 times faster to construct in its Qwen3-8B case study. At 4 bits the result is model-dependent, which the paper explains through weight-outlier structure.

What carries the argument

The load-bearing object is the recursive residual decomposition $\tilde{x}_j^{(t)} = \hat{x}_j^0 + \sum_{k=1}^{t} \hat{r}_j^k$, where $\hat{x}_j^0$ is the dequantized base and $\hat{r}_j^k$ are dequantized 2-bit residual corrections. This identity makes every prefix of stages an additive reconstruction, so higher precision is obtained by summing stage outputs rather than redesigning a bit layout. The paper's analytical support is a two-population outlier model: most weights lie in $[-r,r]$ and one outlier expands a group's range to $[-r,K]$; comparing one-stage versus two-stage quantization gives the threshold $K > r\left(2\cdot \frac{2^B-1}{2^{n_2}-1}-1\right)$, showing residual refinement wins only when localized outliers dominate the group dynamic range. Peak-to-Mean Ratio (maximum absolute weight over mean absolute weight in a group) is introduced as the indicator of such outlier-heavy groups.

What would settle it

Measure the per-group residual radius after the 2-bit RTN base on each evaluated model's weight tensors. If some model's residual max-to-mean ratio stays close to the original group ratio and the 4-bit RRQ prefix underperforms direct RTN 4-bit on that model, the paper's outlier-threshold condition in Section 3.4 is confirmed as the governing mechanism; if residual radii are small yet RRQ still underperforms, the explanation would instead need revision.

Watch

Extended reading notes

Core claim

The central claim is that weight quantization error can itself be quantized, repeatedly, without joint optimization. RRQ starts from a 2-bit round-to-nearest base and then quantizes the residual left by that base, again at 2 bits, and repeats three times; the dequantized approximation at prefix $t$ is the additive sum of the base and the first $t$ residual stages. Because each prefix is a complete weight tensor, a single checkpoint is simultaneously a 2-, 4-, 6-, and 8-bit model. Under the evaluation protocol of MatGPTQ, the paper finds the all-RTN package competitive with MatGPTQ at 6 and 8 bits and ahead of it on four of six models by a 0.1-point tie threshold at those precisions, while 4-bit accuracy depends on the model's outlier profile and on the quality of the base stage. On Qwen3-8B, the full package is constructed in 1,293 seconds, 3.3 times faster than the measured MatGPTQ construction.

Load-bearing premise

The framework rests on the assumption that after a coarse 2-bit base pass, the leftover weight residual is narrow enough in range to be usefully re-quantized at 2 bits; if real quantization groups keep many large residual values, later stages stop reducing error.

Editorial extensions

If this is right

  • One serialized package can serve 2-, 4-, 6-, and 8-bit inference; deployment chooses the prefix at load time instead of requiring a separate checkpoint per precision.
  • A previously quantized low-bit checkpoint can be upgraded toward higher precision by computing residuals from its dequantized weights, without retraining or recalibration.
  • Construction cost for multi-precision packages drops to RTN-level passes: the Qwen3-8B case report is 1,293 seconds versus 4,239 seconds for the measured MatGPTQ baseline.
  • At 4 bits, RRQ is not uniformly better than direct fixed-bit quantization; the paper's model-dependence across the six evaluated LLMs is a real caveat to its multi-precision promise.
  • Stage formats are composable in principle, so a stronger base quantizer such as learned rounding can be swapped in for the 2-bit base while keeping the same residual stages.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves open an explicit test: measuring residual radii after each stage on real weight groups. Models whose residual radius stays far above the inlier radius $r$ should show systematically worse RRQ 4-bit accuracy, and could be identified before building the full package.
  • The 3.3 times construction-speed claim is measured on one model and one protocol; a fairer picture would include end-to-end latency with fused kernels, since accumulating stage GEMMs adds compute and the paper does not report optimized inference timings.
  • If stage formats can be heterogeneous, FP8 or mixed-scale residual stages might close the 4-bit gap on flat-distribution models where 2-bit residuals are too coarse; that is an extension beyond the paper's evaluated integer stages.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Recurrent Residual Quantization (RRQ), a post-training framework that represents an LLM's weights as a 2-bit base plus recurrent 2-bit RTN residual stages, yielding 2-, 4-, 6-, and 8-bit operating points from a single checkpoint. The main empirical claim is that the all-RTN package is competitive with MatGPTQ at 6 and 8 bits under MatGPTQ's evaluation protocol, with model-dependent behavior at 4 bits, and that it is 3.3x faster to construct in a Qwen3-8B case study. Section 3 provides an idealized outlier analysis intended to explain when residual refinement helps, and the appendices report ablations, package sizes, timing, and additional model results.

Significance. RRQ addresses a real deployment problem: providing multiple precision operating points without storing multiple independently quantized checkpoints. The construction is calibration-free, and the timing comparison is concrete and measured. The paper is unusually honest about the model-dependent 4-bit behavior and explicitly does not claim to beat the strongest single-precision quantizers at every bit-width. If the empirical results hold, RRQ is a practical and useful contribution to multi-precision PTQ, though the idealized analysis in Section 3 is not yet a validated predictive theory. The paper also deserves credit for transparently documenting package-size overheads and for isolating the effect of the first-stage quantizer through the RRQ (RTN) versus RRQ (sym) comparison.

major comments (4)
  1. [§3.4, §4, Appendix F] The explanatory mechanism rests on the assumption that after the 2-bit base stage the residual lies in [−r, r] (B_r ≈ r), but the paper never measures stage-wise residual radii on any of the six evaluated models. Appendix F reports only tensor-level PMR and group-maximum K/MAE indicators and explicitly states that these are not intended to predict group-wise reconstruction error. Consequently, the threshold condition in Eq. (13) and the recurrent-benefit interpretation are post hoc qualitative explanations rather than validated claims. Please add direct measurements of per-stage residual radii (or a well-defined proxy) across the evaluated models and report how often the conditions leading to the threshold are actually satisfied; otherwise the Section 3 analysis should be clearly labeled as a speculative toy model rather than an explanation of the empirical results.
  2. [§5.2, Appendix F] The 6- and 8-bit claims are compared only against GPTQ and MatGPTQ numbers copied from the prior paper. Appendix F explicitly omits direct 6- and 8-bit RTN or fixed-bit comparisons. Without such comparisons, the statement that "residual-stage construction can achieve competitive higher-bit prefixes" is not isolated from the possibility that any reasonable 6- or 8-bit quantizer is already near-lossless on these models. Please add direct 6- and 8-bit RTN (and ideally fixed-bit) reconstructions under the same protocol, or restrict the claim to "competitive with MatGPTQ" and state that the residual mechanism is not separately confirmed at higher precisions.
  3. [§5.1, Table 5] All main comparisons rely on single evaluation runs, and the GPTQ/MatGPTQ baselines are quoted rather than reproduced, with no error bars, confidence intervals, or significance tests. The paper nevertheless uses a 0.1 Task Avg tie threshold and counts wins beyond it. Several decisive comparisons (e.g., Qwen3-8B at 6 bits and Qwen3-14B at 8 bits) are within 0.1–0.3 points. Please report variance, bootstrap confidence intervals, or at least multiple evaluation runs for the main comparisons; otherwise the "ahead on four of six models" tally is not statistically defensible.
  4. [§5.1, NeurIPS Checklist items 4–5] Reproducibility is not yet satisfied: code is not released, baseline metrics are copied rather than independently reproduced, and the paper's own checklist answers No for open data and code. For a methods paper whose central evidence is empirical, please provide a detailed reproduction appendix with exact commands, package versions, and a public code release, or clearly state which numbers were independently reproduced and which were taken from prior work.
minor comments (5)
  1. [Table 2] The table does not state the number of elements over which the accumulated absolute error is summed, so the quantitative relation between Eq. (12) and the table cannot be checked; please specify the sequence length and whether the reported value is a sum or a mean.
  2. [§3.4, Eqs. (15)–(16)] Equations (15) and (16) are typeset ambiguously: the factor B_{r,n1} and the fraction (2^B − 1)/(2^{n2} − 1) should be shown with explicit parentheses so the intended algebraic form is unambiguous.
  3. [Appendix F] The first paragraph of Appendix F refers to "Section 4 evaluates the 2+2 RRQ prefix against a direct RTN 4-bit baseline," but the empirical comparison appears in Section 5.2 and Appendix I; the cross-reference should be corrected.
  4. [Table 5] Several cells in Table 5 are visually merged or garbled (e.g., "9.7971.86" and "9.38 75.22"), which makes the table hard to read; insert clear column separators.
  5. [Checklist item 3] The checklist answer labels Section 3 as a proof, but the section contains approximate derivations rather than formal theorems with stated assumptions and complete proofs; consider aligning the checklist phrasing with the actual mathematical content.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RRQ's residual construction is deterministic and self-contained; the Section 3.4 threshold is an explicitly idealized model, and the central accuracy/timing claims are direct measurements.

full rationale

RRQ's construction (Algorithm 1; Eqs. 4-8) is a deterministic successive-approximation scheme: each stage quantizes the residual of the previous stage with a fixed 2-bit RTN quantizer; no parameter is fitted to the accuracy results, and no measured accuracy value is fed back into the construction. The 2/4/6/8-bit operating points are literal prefixes of stored stage tensors, so the multi-precision claim is definitional but not circular. The Section 3.4 threshold analysis is explicitly an idealized model ('Assuming the residual lies within [−r, r]'), and Section 4 concedes that this assumption may fail for realistic distributions and calls for empirically estimated residual radii; it is used to explain, not to generate, the measured accuracy tables. The MatGPTQ comparison uses externally published accuracy numbers plus a measured timing run, not a self-citation chain. The only Intel-authored citation (AutoRound [9], used for fake-quantized QDQ evaluation and the SignRoundV2 ablation) is an implementation tool and is not load-bearing for the central all-RTN claim. The paper's own limitations—no direct 6/8-bit RTN baselines in Appendix F, no error bars, no public code—are correctness/reproducibility concerns, not circular reductions. No fitted quantity is renamed as a prediction, and no claimed result reduces by construction to its inputs.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

No new physical entities or fitted constants; the only ad hoc constructs are the idealized outlier model and the residual-radius compactness assumption, which are explicitly labeled as idealized by the authors. The method's empirical claims do not require those constructs to be true, but the explanatory analysis does.

free parameters (3)
  • stage bit split = 2+2+2+2
    Chosen by hand to provide 2/4/6/8 operating points and a usable 2-bit base; not fitted to task accuracy.
  • group size = 128
    Chosen to match MatGPTQ's grouping layout; ablation shows group 64 improves 2-bit but similar at higher bits.
  • symmetric quantization = symmetric (no zero-points)
    Chosen as main configuration; Appendix I shows near parity with asymmetric on average.
assumptions (5)
  • ad hoc to paper Inlier weights lie in [−r, r] and a single outlier expands the range to [−r, K].
    Toy model in Section 3.3 used to derive error formulas; not verified on real weight groups.
  • ad hoc to paper Residual after the coarse first stage lies in [−r, r] (or has radius B_r ≈ r).
    Section 3.4's idealized RRQ regime; the whole outlier threshold derivation depends on this compactness.
  • standard math Uniform quantizer step-size and expected error formulas, e.g., Δ_direct=(K+r)/(2^B−1) and E≈Δ/4.
    Standard uniform quantization error approximation used in Section 3.4.
  • domain assumption Fake-quantized AutoRound QDQ models faithfully represent the quantization being evaluated.
    All RRQ evaluations use fake-quantized QDQ models (Section 5.1); real optimized kernels are not measured.
  • domain assumption Baseline GPTQ and MatGPTQ numbers taken from MatGPTQ are comparable to locally-run RRQ numbers.
    Section 5.1 says baselines are taken from MatGPTQ under the same evaluation setting; exact harness version and execution differences are not controlled.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Recurrent Residual Quantization: A Progressive Multi-Precision Representation for LLMs." pith.science (2026). https://pith.science/paper/JWYASC6G

@misc{pith2026260804048,
  author       = {Pith},
  title        = {Pith review of: Recurrent Residual Quantization: A Progressive Multi-Precision Representation for LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JWYASC6G}},
  note         = {Machine review of arXiv:2608.04048}
}
read the original abstract

Serving large language models (LLMs) under diverse deployment constraints requires flexible trade-offs between accuracy, memory footprint, and throughput. However, conventional quantization methods typically require a separate checkpoint for each target bit-width. We introduce Recurrent Residual Quantization (RRQ), a post-training quantization (PTQ) framework that represents weights as a low-bit quantized base together with a sequence of quantized residual corrections, enabling multiple effective precisions from a single checkpoint. Starting from a 2-bit model obtained via post-training quantization (PTQ) or round-to-nearest (RTN), RRQ progressively adds lightweight 2-bit residuals generated via RTN to construct 4-, 6-, and 8-bit representations. The method is calibration-free and avoids joint multi-bit optimization. In our Qwen3-8B setup, the full all-RTN 2-/4-/6-/8-bit package is constructed in 1,293 seconds, 3.3 times faster than the measured MatGPTQ construction. Experiments on six recent LLMs show competitive accuracy at 6 and 8 bits, with model-dependent behavior at 4 bits. The code will be made publicly available upon publication.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 7 canonical work pages

  1. [1]

    GPT-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  2. [2]

    Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2024

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2024

    arXiv 2024

  3. [3]

    An Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    arXiv 2025

  4. [4]

    Dubey, A

    A. Dubey, A. Jauhri, A. Pandey, A. Kadian, et al. The Llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024. 9

    arXiv 2024

  5. [5]

    Frantar, S

    E. Frantar, S. Ashkboos, T. Hoefler, and D. Alistarh. Gptq: Accurate post-training quantization for generative pre-trained transformers. InICLR, 2023. arXiv:2210.17323

    arXiv 2023

  6. [6]

    J. Lin, J. Tang, H. Tang, S. Yang, W.-M. Chen, W.-C. Wang, G. Xiao, X. Dang, C. Gan, and S. Han. Awq: Activation-aware weight quantization for on-device llm compression and acceleration. InMLSys, 2024. arXiv:2306.00978

    arXiv 2024

  7. [7]

    J. Chee, Y . Cai, V . Kuleshov, and C. De Sa. Quip: 2-bit quantization of large language models with guarantees. InNeurIPS, 2024. arXiv:2307.13304

    arXiv 2024

  8. [8]

    W. Shao, M. Chen, Z. Zhang, P. Xu, L. Zhao, Z. Li, K. Zhang, P. Gao, Y . Qiao, and P. Luo. Omniquant: Omnidirectionally calibrated quantization for large language models. InICLR,

    work page

Show all 33 references
  1. [9]

    Autoround: Advanced quantization algorithm for llms

    Intel. Autoround: Advanced quantization algorithm for llms. GitHub repository, 2026. https: //github.com/intel/auto-round

    2026

  2. [10]

    Zhong, S

    Y . Zhong, S. Liu, J. Chen, J. Hu, Y . Zhu, X. Liu, X. Jin, and H. Zhang. Distserve: Disaggregating prefill and decoding for goodput-optimized large language model serving. InOSDI, 2024. arXiv:2401.09670

    2024 arXiv

  3. [11]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with PagedAttention. InProceedings of the ACM SIGOPS 29th Symposium on Operating Systems ...

    2023

  4. [12]

    P. Nair, P. Datta, J. Dean, P. Jain, and A. Kusupati. Matryoshka quantization.arXiv preprint arXiv:2502.06786, 2025

    2025 arXiv

  5. [13]

    Kleinegger, E

    M. Kleinegger, E. Crncevic, and D. Alistarh. Matgptq: Accurate and efficient post-training matryoshka quantization.arXiv preprint arXiv:2504.20367, 2025

    2025 arXiv

  6. [14]

    S. Kim, C. Hooper, A. Gholami, Z. Dong, X. Li, S. Shen, M. W. Mahoney, and K. Keutzer. SqueezeLLM: Dense-and-sparse quantization. InICML, 2024. arXiv:2306.07629

    2024 arXiv

  7. [15]

    Dettmers, M

    T. Dettmers, M. Lewis, Y . Belkada, and L. Zettlemoyer. Llm.int8(): 8-bit matrix multiplication for transformers at scale. InNeurIPS, 2022. arXiv:2208.07339

    2022 arXiv

  8. [16]

    G. Xiao, J. Lin, M. Seznec, H. Wu, C. Gan, and S. Han. SmoothQuant: Accurate and efficient post-training quantization for large language models. InICML, 2023. arXiv:2211.10438

    2023 arXiv

  9. [17]

    Z. Liu, B. Oguz, C. Zhao, E. Chang, P. Stock, Y . Meber, R. Memisevic, Y . Shi, and R. Krish- namoorthi. Llm-qat: Data-free quantization aware training for large language models. InACL Findings, 2024. arXiv:2305.17888

    2024 arXiv

  10. [18]

    Z. Dong, Z. Yao, A. Gholami, M. W. Mahoney, and K. Keutzer. HAWQ: Hessian aware quantization of neural networks with mixed-precision. InICCV, 2019. arXiv:1905.03696

    2019 arXiv

  11. [19]

    Z. Dong, Z. Yao, Y . Cai, D. Arfeen, A. Gholami, M. W. Mahoney, and K. Keutzer. HAWQ- V2: Hessian aware trace-weighted quantization of neural networks. InNeurIPS, 2020. arXiv:1911.03852

    2020 arXiv

  12. [20]

    C. Lee, J. Jin, T. Kim, H. Kim, and E. Park. OWQ: Outlier-aware weight quantization for efficient fine-tuning and inference of large language models. InAAAI, 2024. arXiv:2306.02272

    2024 arXiv

  13. [21]

    Juang and A

    B.-H. Juang and A. H. Gray. Multiple stage vector quantization for speech coding. InICASSP, 1982

    1982

  14. [22]

    Martinez, S

    J. Martinez, S. Zakhmi, H. H. Hoos, and N. de Vries. Permute, quantize, and fine-tune: Efficient compression of neural networks. InCVPR, 2021. arXiv:2012.09659

    2021 arXiv

  15. [23]

    J. Lee, S. Cho, and S.-K. Beack. Context-adaptive residual coding for image compression. In CVPR, 2022. arXiv:2203.08862. 10

    2022 arXiv

  16. [24]

    H. Xi, S. Yang, Y . Zhao, M. Li, H. Cai, X. Li, Y . Lin, Z. Zhang, J. Zhang, X. Li, Z. Xu, J. Wu, C. Xu, I. Stoica, S. Han, and K. Keutzer. Quant videogen: Auto-regressive long video generation via 2-bit kv-cache quantization.arXiv preprint arXiv:2602.02958, 2026

    2026 arXiv

  17. [25]

    M. Sun, X. Chen, J. Z. Kolter, and Z. Liu. Massive activations in large language models. In COLM, 2024. arXiv:2402.17762

    2024 arXiv

  18. [26]

    Dettmers, R

    T. Dettmers, R. Svirschevski, V . Egiazarian, D. Kuznetsov, E. Frantar, and D. Alistarh. Spqr: A sparse-quantized representation for near-lossless llm weight compression. InICLR, 2024. arXiv:2306.03078

    2024 arXiv

  19. [27]

    Abdin, S

    M. Abdin, S. A. Jacobs, A. A. Awan, J. Aneja, A. Awadallah, et al. Phi-3 technical report: A highly capable language model locally on your phone.arXiv preprint arXiv:2404.14219, 2024

    2024 arXiv

  20. [28]

    L. Gao, J. Tow, B. Abbasi, S. Biderman, S. Black, A. DiPofi, C. Foster, L. Golding, J. Hsu, A. Le Noac’h, H. Li, K. McDonell, N. Muennighoff, C. Ociepa, J. Phang, L. Reynolds, H. Schoelkopf, A. Skowron, L. Sutawika, E. Tang, A. Thite, B. Wang, K. Wang, and A. Zou. The language...

    2024

  21. [29]

    Clark, I

    P. Clark, I. Cowhey, O. Etzioni, T. Khot, A. Sabharwal, C. Schoenick, and O. Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge.arXiv preprint arXiv:1803.05457, 2018

    2018 arXiv

  22. [30]

    Zellers, A

    R. Zellers, A. Holtzman, Y . Bisk, A. Farhadi, and Y . Choi. Hellaswag: Can a machine really finish your sentence? InACL, 2019. arXiv:1905.07830

    2019 arXiv

  23. [31]

    Y . Bisk, R. Zellers, R. LeBras, J. Gao, and Y . Choi. Piqa: Reasoning about physical common- sense in natural language. InAAAI, 2020. arXiv:1911.11641

    2020 arXiv

  24. [32]

    Separate est

    K. Sakaguchi, R. L. Bras, C. Bhagavatula, and Y . Choi. Winogrande: An adversarial winograd schema challenge at scale.Communications of the ACM, 64(9):99–106, 2021. arXiv:1907.10641. A Broader Impacts and Asset Licenses Broader impacts.RRQ may improve deployment efficiency and...

    2021 arXiv

  25. [33]

    Guidelines: • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.