Pith. sign in

REVIEW 4 major objections 6 minor 46 references

Loss-aware rank budgets and residual-stream corrections keep SVD-compressed LLMs far closer to the original model at high compression.

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 · grok-4.5

2026-07-12 05:10 UTC pith:S773IVGN

load-bearing objection Solid post-training SVD recipe with real high-ratio gains; the main soft spot is WikiText-2 calibration/eval overlap, not a broken method. the 4 major comments →

arxiv 2607.03057 v1 pith:S773IVGN submitted 2026-07-03 cs.LG cs.AI

LACE-SVD: Loss-Aware SVD with Cumulative Error Correction for LLM Compression

Zhuowen Liu , Longkun Hao , Shiyu Feng , Xiaowen Chang , Ruiqun Li , Changqun Li This is my paper
classification cs.LG cs.AI
keywords LLM compressionsingular value decompositionlow-rank factorizationloss-aware rank allocationcumulative error correctionresidual streampost-training compressioncalibration NLL
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.

Existing SVD compressors for large language models mostly minimize local matrix reconstruction error. That leaves two gaps: rank is often spent without asking which layers hurt language modeling most, and small errors in residual-stream projections accumulate across the stack. LACE-SVD closes both. It scores candidate keep-ratios by the rise in calibration negative log-likelihood they cause when a single layer is compressed, then solves a budgeted allocation so loss-sensitive layers keep more capacity. After factorization it applies closed-form local least-squares updates and a gated, propagation-aware correction only on the residual-stream output modules (attention out and MLP down). On LLaMA-7B at 60 percent compression the result is WikiText-2 perplexity 32.57 versus 46-plus for the strongest prior SVD baselines, with matching gains on other models and memory budgets. A sympathetic reader cares because the method stays post-training, hardware-agnostic, and still delivers usable language models under aggressive size and latency targets.

Core claim

The paper shows that coupling calibration-NLL-based layer-wise rank allocation with closed-form local updates and a gated residual-stream output correction produces substantially higher end-to-end fidelity than purely local SVD compressors, especially at high compression ratios, without end-to-end fine-tuning.

What carries the argument

LACE-SVD: loss-aware multiple-choice knapsack allocation of keep-ratios (via dynamic programming on per-layer calibration-NLL deltas) followed by simultaneous closed-form local updates and a propagation-aware target for residual-stream modules, accepted only when layer-output L2 fidelity improves.

Load-bearing premise

That scoring each layer alone with calibration loss, then correcting residual-stream modules by layer-output distance rather than the true language-modeling loss, is still enough to choose ranks and fixes that hold up end-to-end.

What would settle it

Re-run the same LLaMA-7B 0.6 protocol but replace the independent NLL probes with a joint multi-layer search (or replace the L2 residual gate with direct LM-loss acceptance); if WikiText-2 PPL and zero-shot averages collapse back toward the SVD-LLM / Dobi-SVD numbers, the proxy chain fails.

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

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 / 6 minor

Summary. LACE-SVD is a post-training SVD compression method for LLMs that (i) allocates layer-wise keep-ratios by measuring calibration negative-log-likelihood increases for candidate ratios and solving a budgeted multiple-choice allocation (Eqs. 5–8, DP), and (ii) after activation-whitened SVD, applies simultaneous closed-form local factor updates plus a propagation-aware cumulative error correction (CEC) on residual-stream modules (o_proj, down_proj), with target T_pa (Eq. 14) and a layer-output L2 acceptance gate (Eq. 15). On LLaMA-7B at compression ratio 0.6, the paper reports WikiText-2 PPL 32.57 vs Dobi-SVD 46.18 and SVD-LLM 53.74 (Table 1), with ablations attributing gains to allocation, local update, and CEC (Table 4), plus multi-architecture results (Table 2), a pruning comparison under matched memory (Table 3), and a consistent U-shaped rank pattern (Fig. 2).

Significance. If the gains hold under clean evaluation, the work is a useful systems contribution: it makes rank allocation explicitly loss-sensitive under a global budget and adds a practical residual-stream correction without end-to-end fine-tuning, with clear empirical separation of components (Table 4) and multi-model evidence (LLaMA/OPT/Mistral/Vicuna). Strengths include closed-form local updates, an acceptance gate that can reject CEC, matched-memory pruning comparisons, and explicit Limitations on independent-layer probes and L2 proxies. The high-ratio regime (0.6–0.8) is where prior SVD methods degrade sharply, so a robust fix would matter for deployment.

major comments (4)
  1. Primary evaluation and calibration share WikiText-2. Appendix B/D state that whitening, loss-aware candidate evaluation, local updates, and the CEC acceptance gate all use WikiText-2 samples, while Table 1’s headline metric and Table 4’s full ablation (allocation / local update / CEC) report WikiText-2 PPL. At ratio 0.6 the central claim (32.57 vs 46.18/53.74) and the component attributions are therefore not cleanly separated from calibration–test overlap. C4 and zero-shot help, but they do not replace: (a) ablations on a held-out corpus (e.g., C4 PPL for each Table 4 row), and (b) a protocol where allocation/CEC statistics are built only on data disjoint from the reported PPL corpus (or a second calibration source). Without this, the load-bearing high-ratio result remains under-supported.
  2. §3.3–3.5 and Limitations treat independent single-layer ΔNLL probes (Eq. 5) and residual-stream L2 (Eqs. 14–15 on o_proj/down_proj only) as proxies that ignore nonlinear cross-layer coupling and never optimize LM loss. That is acceptable as a design choice, but the paper’s claim that these proxies preserve end-to-end fidelity under a global budget needs a direct stress test: e.g., (i) sequential vs simultaneous allocation (compress layers in order and re-measure ΔNLL), (ii) CEC applied to a broader module set vs S only, and (iii) correlation between accepted layer-L2 reductions and held-out NLL/zero-shot. Currently Table 4 shows additive gains on WikiText-2 only; that is insufficient to validate the proxy axioms for the global claim.
  3. Related work cites SAES-SVD (Hu et al., 2026) as also targeting cross-layer error accumulation with a cumulative error-aware objective, yet experiments never compare to it (or to SVD-LLM v2’s theoretical-loss layer-wise ratios). For the novelty claim that LACE-SVD specifically fixes “heuristic rank allocation and cumulative error propagation,” a head-to-head under the same calibration budget and ratio 0.6 on LLaMA-7B is load-bearing; otherwise the contribution relative to concurrent cumulative-error SVD work is unclear.
  4. No uncertainty quantification. All tables report single-run PPL/accuracy with no seeds, calibration-subset resampling, or error bars. At aggressive ratios (Table 1, 0.6–0.8; Table 4), absolute PPL differences of several points could be sensitive to the 256-sample WikiText-2 draw and free parameters (α, λ_U/λ_V, R_ρ, DP bins; Appendix B/C). At least multi-seed or multi-calibration-split results for the 0.6 LLaMA-7B setting and the Table 4 ablations are needed before the ranking over Dobi-SVD/SVD-LLM can be treated as stable.
minor comments (6)
  1. Figure 1’s text is hard to read in places (garbled symbols in the pipeline diagram); please regenerate with legible math for whitening and CEC.
  2. Appendix A Algorithm 1 is useful; cross-reference it earlier in §3.1 and ensure notation matches the main text (cW vs Ŵ, r⋆_ℓ).
  3. Table 1 marks some baselines with † (fine-tuning) and ASVD∗; clarify in the caption which methods are pure post-training under identical no-FT protocol so the 0.6 comparison is apples-to-apples.
  4. α ablation (Appendix C, Fig. 5) is informative but only on WikiText-2 at 0.6; a short note on whether α=0.7 remains best on C4 would strengthen the hyperparameter claim.
  5. Typos/clarity: “LACE-SVD first estimates the calibration negative-log-likelihood increase” is fine; ensure consistent “keep-ratio” vs “compression ratio” (ρ=0.6 as keep vs remove is easy to misread—state explicitly).
  6. Fig. 2 U-shaped allocation is a nice empirical finding; consider quantifying early/late vs middle keep-ratios (mean r per band) in the text for reproducibility.

Circularity Check

1 steps flagged

Empirical post-training compressor; no load-bearing derivation reduces to its inputs by construction—only mild hyperparameter selection on the reported WikiText-2 metric.

specific steps
  1. fitted input called prediction [Appendix C / Figure 5; main claim Table 1 (ratio 0.6)]
    "the framework achieves the best perplexity (32.57) when α=0.7 ... Based on these empirical findings, we fix α=0.7 for all our main experiments ... Experimental results demonstrate that at a high compression ratio (0.6), the WikiText-2 PPL of our method on LLaMA-7B (32.57)"

    Correction strength α is selected by sweeping values and retaining the setting that minimizes WikiText-2 PPL at compression ratio 0.6; the same corpus, ratio, and PPL number are then reported as the headline result. The reported 32.57 is therefore partly the outcome of choosing the hyperparameter that optimizes that metric, not an independent out-of-sample prediction of α. This is mild (one scalar, ablated openly) and does not make allocation or CEC tautological.

full rationale

LACE-SVD is a systems/methods paper, not a first-principles derivation. Rank allocation (Eq. 5–8) minimizes measured single-layer calibration ΔNLL under a budget via DP; CEC (Eq. 14–15) re-solves residual-stream factors toward an L2 layer-output proxy with an acceptance gate. Neither step defines the headline WikiText-2/C4/zero-shot numbers by construction: full multi-layer compression, simultaneous local updates, and the gate can still fail or transfer poorly, and C4/zero-shot provide external checks. Related-work citations (ASVD, SVD-LLM, Dobi-SVD, SAES-SVD) are prior methods by other groups used as baselines or context, not uniqueness theorems that force the present design. The only mild circularity-adjacent practice is choosing α=0.7 by ablating WikiText-2 PPL at the same 0.6 ratio used for the main claim; that is ordinary hyperparameter tuning on the reported metric, not a self-definitional or self-citation chain. Calibration–eval corpus overlap (WikiText-2) is a validity/leakage concern, not circularity of the derivation. Score 1 reflects that minor fitted-input issue only.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central claim rests on standard linear-algebra SVD practice plus domain modeling choices: that single-layer calibration NLL deltas rank layer importance under a global budget, that residual-stream writers dominate cumulative error, and that L2 layer-output fidelity is a usable proxy without end-to-end LM loss optimization. Several scalars (α, ridge terms, gate thresholds, candidate grids) are chosen by hand or ablation rather than derived. No new physical entities; invented pieces are algorithmic constructs (allocation DP objective, CEC target, acceptance gate).

free parameters (5)
  • correction strength α = 0.7
    Blends full-precision and compressed residual-module outputs in Eq. 14; fixed to 0.7 after WikiText-2 ablation at ratio 0.6 (Fig. 5 / Appendix C).
  • ridge coefficients λU, λV = λU=1e-5, λV=1e-4
    Regularize closed-form local factor updates (Eq. 12–13); set to 1e-5 and 1e-4 in Appendix D.3 without derivation from first principles.
  • minimum held-out relative gain = 2e-4
    Acceptance threshold for optional bi-side updates; set to 2e-4 (Table 5 / D.3).
  • candidate keep-ratio set Rρ and DP budget bins = bins=4000; Rρ target-dependent
    Discretize the multiple-choice knapsack; treated as experimental hyperparameters (D.1–D.2), bins=4000.
  • calibration sample counts / sequence lengths = 256 / 64 / 64 as in Table 5
    256 whitening samples @2048; 64 loss-aware batches @1024; local-update 64 samples—chosen for cost/accuracy tradeoff, not derived.
axioms (4)
  • domain assumption Activation-whitened truncated SVD is a valid local compressor for LLM linear maps (G≈XXᵀ, WC=UΣVᵀ).
    Inherited from ASVD/SVD-LLM line; §3.2 treats it as the base factorization path.
  • ad hoc to paper Sum of independent single-layer calibration NLL increases is a usable objective for global rank allocation under a parameter budget.
    Eq. 5–8 and Limitations admit ignored cross-layer coupling; tractability motivates the DP formulation.
  • ad hoc to paper Layer-output L2 discrepancy on residual-stream modules is a practical proxy for cumulative error affecting end-to-end behavior.
    §3.5 explicitly does not optimize LM loss; gate Eq. 15 retains CEC only if layer L2 improves.
  • domain assumption Post-training closed-form least-squares updates on calibration activations improve factors without full fine-tuning.
    Standard reconstruction assumption in post-training compression; §3.4 simultaneous mode uses full-precision activations.
invented entities (2)
  • Propagation-aware CEC target T_pa with acceptance gate no independent evidence
    purpose: Nudge residual-stream module outputs toward full-precision contributions and keep the update only if whole-layer L2 improves.
    Algorithmic construct (Eq. 14–15); independent_evidence false because utility is only shown via the paper’s own compressed-model metrics.
  • Loss-aware multiple-choice DP rank allocator on measured ΔNLL tables no independent evidence
    purpose: Assign heterogeneous keep-ratios under global compression budget ρ.
    Optimization wrapper around calibration probes; related to prior adaptive-rank ideas but specified as this paper’s allocation stage.

pith-pipeline@v1.1.0-grok45 · 22319 in / 3809 out tokens · 35826 ms · 2026-07-12T05:10:12.770340+00:00 · methodology

0 comments
read the original abstract

The rapid growth in the parameter scale of large language models (LLMs) has created a strong demand for efficient compression techniques. As a hardware-agnostic and highly compatible approach, low-rank compression has been widely adopted to reduce both memory footprint and computational cost. However, existing SVD-based methods are still largely driven by local reconstruction objectives, overlooking two critical limitations: rank budgets are often allocated without explicitly considering layer-wise loss sensitivity, and local approximation errors can propagate and accumulate through the residual stream, leading to amplified global deviations from the original model. To address these issues, we propose LACE-SVD, a Loss-Aware SVD framework with Cumulative Error correction for LLM compression. LACE-SVD first estimates the calibration negative-log-likelihood increase induced by candidate layer-wise compression ratios and solves a budget-constrained allocation problem to assign rank budgets. It then refines the compressed model with closed-form local updates and introduces a propagation-aware correction for residual-stream output modules, reducing layer-output discrepancy as a proxy for cumulative error propagation. Experimental results demonstrate that at a high compression ratio (0.6), the WikiText-2 PPL of our method on LLaMA-7B (32.57) is significantly better than that of Dobi-SVD (46.18).

Figures

Figures reproduced from arXiv: 2607.03057 by Changqun Li, Longkun Hao, Ruiqun Li, Shiyu Feng, Xiaowen Chang, Zhuowen Liu.

Figure 1
Figure 1. Figure 1: Overall architecture of LACE-SVD. Original LLM layers are first decomposed into initial low-rank weights using Activation-Whitened SVD. A loss-aware allocation strategy then dynamically selects the optimal rank ratio (rg) for each layer. Finally, an error correction pipeline, incorporating Local Updates and Cumulative Error Correction (CEC), is applied to compensate for the cumulative errors before assembl… view at source ↗
Figure 2
Figure 2. Figure 2: Layer-wise compression ratio allocation for LLaMA-7B under different targets. The loss-aware allocation strategy dynamically assigns compression budgets across 32 layers. Compared to the uniform baseline (red dashed line), the algorithm consistently preserves more parameters in the early layers (L00–L06, green region) to maintain foundational lexical and syntactic features, as well as in the final layers (… view at source ↗
Figure 3
Figure 3. Figure 3: Perplexity (↓) and average accuracy (↑) of LLaMA-13B under 20% compression. 0.0 0.5 1.0 1.5 2.0 SpeedUp 0% 20% 40% 60% Compression Ratio 12.90GB 1.00x 10.55GB 1.26x 8.15GB 1.51x 5.60GB 1.97x Baseline 20% Compression 40% Compression 60% Compression [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Memory usage and inference speedup of LLaMA-7B on varying compression ratios [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Ablation study of the correction strength [PITH_FULL_IMAGE:figures/full_fig_p012_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

46 extracted references · 1 canonical work pages

  1. [1]

    SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression , author=

  2. [2]

    SVD - LLM V2: Optimizing Singular Value Truncation for Large Language Model Compression

    Wang, Xin and Alam, Samiul and Wan, Zhongwei and Shen, Hui and Zhang, Mi. SVD - LLM V2: Optimizing Singular Value Truncation for Large Language Model Compression. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025. doi:10.1865...

    work page doi:10.18653/v1/2025.naacl-long.217 2025

  3. [3]

    Language model compression with weighted low-rank factorization , author=

  4. [4]

    arXiv preprint arXiv:2312.05821 , year=

    Asvd: Activation-aware singular value decomposition for compressing large language models , author=. arXiv preprint arXiv:2312.05821 , year=

    Pith/arXiv arXiv

  5. [5]

    arXiv preprint arXiv:2502.02723 , year=

    Dobi-SVD: Differentiable SVD for LLM Compression and Some New Perspectives , author=. arXiv preprint arXiv:2502.02723 , year=

    Pith/arXiv arXiv

  6. [6]

    Geophysical Prospecting , volume=

    Dip-adaptive singular-value decomposition filtering for seismic reflection enhancement , author=. Geophysical Prospecting , volume=. 2013 , publisher=

    2013

  7. [7]

    arXiv preprint arXiv:2602.03051 , year=

    SAES-SVD: Self-Adaptive Suppression of Accumulated and Local Errors for SVD-based LLM Compression , author=. arXiv preprint arXiv:2602.03051 , year=

    arXiv

  8. [8]

    Proceedings of the 37th International Conference on Neural Information Processing Systems , articleno =

    Ma, Xinyin and Fang, Gongfan and Wang, Xinchao , title =. Proceedings of the 37th International Conference on Neural Information Processing Systems , articleno =. 2024 , publisher =

    2024

  9. [9]

    Croci and Marcelo Gennari do Nascimento and Torsten Hoefler and James Hensman , booktitle=

    Saleh Ashkboos and Maximilian L. Croci and Marcelo Gennari do Nascimento and Torsten Hoefler and James Hensman , booktitle=. Slice. 2024 , url=

    2024

  10. [10]

    12th International Conference on Learning Representations, ICLR 2024 , year=

    A SIMPLE AND EFFECTIVE PRUNING APPROACH FOR LARGE LANGUAGE MODELS , author=. 12th International Conference on Learning Representations, ICLR 2024 , year=

    2024

  11. [11]

    arXiv preprint arXiv:2402.05406 , year=

    Everybody prune now: Structured pruning of llms with only forward passes , author=. arXiv preprint arXiv:2402.05406 , year=

    arXiv

  12. [12]

    Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining , pages=

    Model compression , author=. Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining , pages=

  13. [13]

    arXiv preprint arXiv:1710.09282 , year=

    A survey of model compression and acceleration for deep neural networks , author=. arXiv preprint arXiv:1710.09282 , year=

    Pith/arXiv arXiv

  14. [14]

    Artificial Intelligence Review , volume=

    A comprehensive survey on model compression and acceleration , author=. Artificial Intelligence Review , volume=. 2020 , publisher=

    2020

  15. [15]

    Advances in neural information processing systems , volume=

    Qlora: Efficient finetuning of quantized llms , author=. Advances in neural information processing systems , volume=

  16. [16]

    Proceedings of machine learning and systems , volume=

    Awq: Activation-aware weight quantization for on-device llm compression and acceleration , author=. Proceedings of machine learning and systems , volume=

  17. [17]

    arXiv preprint arXiv:2303.08774 , year=

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    Pith/arXiv arXiv

  18. [18]

    int8 (): 8-bit matrix multiplication for transformers at scale , author=

    Gpt3. int8 (): 8-bit matrix multiplication for transformers at scale , author=. Advances in neural information processing systems , volume=

  19. [19]

    arXiv preprint arXiv:2302.13971 , year=

    Llama: Open and efficient foundation language models , author=. arXiv preprint arXiv:2302.13971 , year=

    Pith/arXiv arXiv

  20. [20]

    arXiv preprint arXiv:2407.21783 , year=

    The llama 3 herd of models , author=. arXiv preprint arXiv:2407.21783 , year=

    Pith/arXiv arXiv

  21. [21]

    arXiv preprint arXiv:2307.09288 , year=

    Llama 2: Open foundation and fine-tuned chat models , author=. arXiv preprint arXiv:2307.09288 , year=

    Pith/arXiv arXiv

  22. [22]

    arXiv preprint arXiv:2309.16609 , year=

    Qwen technical report , author=. arXiv preprint arXiv:2309.16609 , year=

    Pith/arXiv arXiv

  23. [23]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

  24. [24]

    arXiv preprint arXiv:2210.17323 , year=

    Gptq: Accurate post-training quantization for generative pre-trained transformers , author=. arXiv preprint arXiv:2210.17323 , year=

    Pith/arXiv arXiv

  25. [25]

    arXiv preprint arXiv:2405.17849 , year=

    I-llm: Efficient integer-only inference for fully-quantized low-bit large language models , author=. arXiv preprint arXiv:2405.17849 , year=

    Pith/arXiv arXiv

  26. [26]

    arXiv preprint arXiv:1503.02531 , year=

    Distilling the knowledge in a neural network , author=. arXiv preprint arXiv:1503.02531 , year=

    Pith/arXiv arXiv

  27. [27]

    Psychometrika , volume=

    The approximation of one matrix by another of lower rank , author=. Psychometrika , volume=. 1936 , publisher=

    1936

  28. [28]

    Linear Algebra and its Applications , volume=

    A generalization of the Eckart-Young-Mirsky matrix approximation theorem , author=. Linear Algebra and its Applications , volume=. 1987 , publisher=

    1987

  29. [30]

    The Thirteenth International Conference on Learning Representations , year=

    OSTQuant: Refining Large Language Model Quantization with Orthogonal and Scaling Transformations for Better Distribution Fitting , author=. The Thirteenth International Conference on Learning Representations , year=

  30. [31]

    arXiv preprint arXiv:2106.08295 , year=

    Loraprune: Pruning meets low-rank parameter-efficient fine-tuning , author=. arXiv preprint arXiv:2106.08295 , year=

    Pith/arXiv arXiv

  31. [32]

    Advances in neural information processing systems , volume=

    Llm-pruner: On the structural pruning of large language models , author=. Advances in neural information processing systems , volume=

  32. [33]

    12th International Conference on Learning Representations , year=

    SpQR: A sparse-quantized representation for near-lossless LLM weight compression , author=. 12th International Conference on Learning Representations , year=

  33. [34]

    arXiv preprint arXiv:2502.01403 , year=

    Adasvd: Adaptive singular value decomposition for large language models , author=. arXiv preprint arXiv:2502.01403 , year=

    arXiv

  34. [35]

    International conference on machine learning , pages=

    Optimizing neural networks with kronecker-factored approximate curvature , author=. International conference on machine learning , pages=. 2015 , organization=

    2015

  35. [36]

    arXiv e-prints , pages=

    Pointer Sentinel Mixture Models , author=. arXiv e-prints , pages=

  36. [37]

    Journal of machine learning research , volume=

    Exploring the limits of transfer learning with a unified text-to-text transformer , author=. Journal of machine learning research , volume=

  37. [38]

    CoRR , volume =

    Peter Clark and Isaac Cowhey and Oren Etzioni and Tushar Khot and Ashish Sabharwal and Carissa Schoenick and Oyvind Tafjord , title =. CoRR , volume =. 2018 , url =. 1803.05457 , timestamp =

    Pith/arXiv arXiv 2018

  38. [39]

    H ella S wag: Can a Machine Really Finish Your Sentence?

    Zellers, Rowan and Holtzman, Ari and Bisk, Yonatan and Farhadi, Ali and Choi, Yejin. H ella S wag: Can a Machine Really Finish Your Sentence?. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019. doi:10.18653/v1/P19-1472

    doi:10.18653/v1/p19-1472 2019

  39. [40]

    M ath QA : Towards Interpretable Math Word Problem Solving with Operation-Based Formalisms

    Amini, Aida and Gabriel, Saadia and Lin, Shanchuan and Koncel-Kedziorski, Rik and Choi, Yejin and Hajishirzi, Hannaneh. M ath QA : Towards Interpretable Math Word Problem Solving with Operation-Based Formalisms. Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, ...

    doi:10.18653/v1/n19-1245 2019

  40. [41]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Piqa: Reasoning about physical commonsense in natural language , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  41. [42]

    arXiv preprint arXiv:1907.10641 , year=

    WinoGrande: An Adversarial Winograd Schema Challenge at Scale , author=. arXiv preprint arXiv:1907.10641 , year=

    Pith/arXiv arXiv 1907

  42. [43]

    2023 , publisher=

    Stanford alpaca: An instruction-following llama model , author=. 2023 , publisher=

    2023

  43. [44]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    Blockpruner: Fine-grained pruning for large language models , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  44. [45]

    arXiv preprint arXiv:2205.01068 , year=

    Opt: Open pre-trained transformer language models , author=. arXiv preprint arXiv:2205.01068 , year=

    Pith/arXiv arXiv

  45. [46]

    See https://vicuna

    Vicuna: An open-source chatbot impressing gpt-4 with 90\ author=. See https://vicuna. lmsys. org (accessed 14 April 2023) , volume=

    2023

  46. [47]

    2023 , eprint=

    Mistral 7B , author=. 2023 , eprint=

    2023