Pith. sign in

REVIEW 3 major objections 6 minor 23 references

Aligned language models often prefer the wrong answer in mid layers and only rescue correctness late—hiding fragility that compression and surface-only fine-tuning can expose.

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

2026-07-11 15:57 UTC pith:IBP2FU7P

load-bearing objection Careful multi-model causal story of late rescue: item-level wrong-dip predicts structural compression risk, is recipe-specific, and can be trained down—scope is controlled instruments, not open dialogue. the 3 major comments →

arxiv 2607.04640 v1 pith:IBP2FU7P submitted 2026-07-06 cs.CL

Wrong Before Right: Late Rescue and Interface Failure in Aligned Language Models

Jiaqi Deng This is my paper
classification cs.CL
keywords wrong-diplate rescuealigned language modelsactivation transplantationmodel compressioninternal trajectoriesdifference-in-differencespost-training
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper claims that final-answer correctness in aligned language models can hide how that correctness is assembled inside the network. On polarity-controlled minimal pairs, mid layers (roughly 25–90% depth) often transiently prefer the incorrect answer—the wrong-dip—before late layers correct it; the effect is causally verified by transplanting mid-layer activations into a neutral decoding context. The dip is amplified by some alignment recipes in a scale-emergent, family-specific way while remaining invisible at the output. High-dip items are far more likely to flip under genuine late structural compression, block dropping, or pruning, but not under quantization; a mid-layer wrong-margin penalty during LoRA fine-tuning can cut the causal dip by about two-thirds at matched accuracy and improve compression robustness, whereas output-only SFT can worsen the internal dip. With controlled readouts the pattern reaches naturalistic vignettes and splits free-form fragility into a dip-auditable late-rescue layer and a dip-blind interface layer. A sympathetic reader cares because output-level safety and accuracy tests can certify models whose correct behavior is one compression step or one interface failure away from collapsing.

Core claim

Correctness in aligned language models is often assembled by late rescue: mid-layer internal preference transiently commits to the wrong answer (the wrong-dip), and only late layers correct it. The phenomenon is measured by layer-wise difference-in-differences trajectories on polarity-controlled minimal pairs, verified causally by patchscope-style activation transplantation across 17 models spanning three families and 0.5B–32B, and shown to govern structural compression risk, post-training internal quality, and a separable class of natural-language evaluation failures while remaining invisible to output accuracy.

What carries the argument

The wrong-dip, measured item-wise as max_wrong_dip (deepest wrong-direction excursion of the DiD trajectory in the 25–90% depth window) and commit_layer_frac (last relative depth whose internal preference still disagrees with the final output), and verified by transplanting mid-layer hidden states into a neutral decoding context so the model produces the wrong token.

Load-bearing premise

The load-bearing premise is that polarity-controlled minimal pairs scored by layer-wise difference-in-differences, plus mid-layer activation transplants into a neutral context, are a faithful enough proxy for how models assemble correctness on the relation-assembly tasks that matter in real deployment.

What would settle it

If items with large intact-model wrong-dips do not flip at substantially higher rates than low-dip items under genuine late-layer structural damage (low-rank compression, block dropping, or structured pruning) on held-out natural vignettes with semantic-candidate readouts—while the same items remain unpredictive under quantization—the late-rescue account would fail.

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

If this is right

  • High-dip items on the intact model are 3–7× more likely to flip under late-layer low-rank compression, block dropping, or structured pruning, while quantization flips are dip-blind.
  • Output-only supervised fine-tuning can reach perfect surface accuracy while worsening the causal internal dip by up to 2.8×, increasing hidden structural risk.
  • A LoRA fine-tune that adds a mid-layer wrong-margin hinge penalty can cut the causal dip 67–70% at matched accuracy and transfer mid-layer compression robustness.
  • Alignment amplification of the causal dip is recipe-specific and emergent (appears at 3B and peaks at 32B in Qwen2.5; reverses in Llama-3-8B), so the dip audits recipes rather than alignment itself.
  • Once readouts are controlled, free-form evaluation fragility separates into a dip-auditable late-rescue failure and a dip-blind interface-binding failure.

Where Pith is reading between the lines

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

  • Safety and capability certifications that look only at final answers may systematically overstate reliability for models that will later be pruned, low-rank compressed, or further fine-tuned.
  • Family differences in how correctness is delegated (late-layer patch versus mid-layer rewrite) may help explain known gaps in low-bit quantization survival without separate ad-hoc mechanisms for each failure mode.
  • Tracking the wrong-dip after output accuracy saturates could become a practical post-training monitor that flags recipes polishing only the surface before any compression test is run.
  • If relational assembly load drives the dip, multi-binding tasks may be the primary hidden failure surface under structural damage even when single-token accuracy looks intact.

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

3 major / 6 minor

Summary. The paper argues that output-level correctness in aligned LMs can hide a late-rescue process: on polarity-controlled minimal pairs, mid-layer DiD trajectories (25–90% depth) transiently prefer the incorrect answer (the wrong-dip), and late layers correct it. The dip is verified causally by patchscope-style mid-layer activation transplantation across 17 models (three families, 0.5B–32B). Four claims follow: (1) alignment amplification of the causal dip is recipe-specific and emergent (emerges at 3B in Qwen2.5, peaks at 32B; reverses in Llama-3-8B); (2) intact-model dip predicts item flips under late-layer SVD, block dropping, and structured pruning (3–7×) but not under quantization—a double dissociation confirmed by late residual attenuation; (3) a mid-layer wrong-margin LoRA penalty matches output-only SFT accuracy while cutting the causal dip 67–70% and improving mid-SVD retention, whereas output-only SFT can worsen the causal dip up to 2.8×; (4) with semantic-candidate readouts the phenomenon extends to naturalistic vignettes, and free-form fragility separates into a dip-auditable late-rescue layer and a dip-blind interface layer.

Significance. If the results hold, the paper supplies a concrete, item-level account of how correctness is assembled rather than merely whether the final answer is right, with direct consequences for structural compression risk, post-training quality, and evaluation design. Strengths that raise the contribution above a pure observational finding include: causal patchscopes on the full n=278 set across the scale ladder; de-circularized prediction (dip always measured on the intact model); causal re-measurement of all trained arms; multi-seed McNemar significance for 7B robustness transfer; and an explicit structural-vs-quantization double dissociation with late residual deprivation. The recipe taxonomy and training-time monitor are practically useful. External validity to fully open-ended dialogue remains a stated scope boundary, but within the controlled instruments the evidence is unusually thorough for an internal-mechanism paper.

major comments (3)
  1. §5.6 Acts 1–4 and Limitations: free-generation mechanism stratification is left open, and free-form failure is shown to be cliff-like and family-dependent under the damage titration. Finding (4) and the abstract claim that free-form fragility separates into a dip-auditable late-rescue layer and a dip-blind interface layer therefore rest primarily on the semantic-candidate (L1) and constrained (L2) rungs of the bridge ladder, not on open generation. The paper should either (a) report a successful mild-damage free-generation stratification on the graded Llama regime, or (b) narrow the abstract/§8 claim so that the interface split is stated only for controlled semantic/constrained readouts, with free-generation listed as feasibility-only.
  2. §5.7 and Limitations: the relational-complexity ladder’s long-distance tier was retired post hoc as a design flaw (redundant cue near the query). The upgrade from a “negation finding” to a “relation-assembly under late correction budget” claim therefore rests on the remaining three tiers plus the agent/patient pairs. A short sensitivity check—re-running the two-relation peak after removing any residual surface confounds, or replacing the retired tier with a clean long-distance construction—would make the core-mechanism claim load-bearing rather than provisional.
  3. §6.3–6.5 and Table 4: multi-seed significance and the dip→retention forecast are established for mid-SVD r64 on Qwen2.5-7B (and partially on 1.5B), but late-SVD and block-drop transfer of the regularizer are not reported with the same seed structure, and Llama-8B ceilings the retention scale. Because the audit’s strongest predictive signal is late-layer structural damage (§5.2–5.3), the intervention’s claim that “the fix transfers to structural-compression robustness” should either include late-SVD/block-drop McNemar results or be scoped explicitly to mid-layer structural operators.
minor comments (6)
  1. §3.3: the raw-logit-lens overestimate for Llama (0.72 vs causal 0.023) is an important methodological warning; consider elevating it to a short boxed note or table so cross-family readers do not miss it.
  2. Figure 3 / Figure 5: axis labels and depth-window shading for the 25–90% band would make the wrong-dip visually self-explanatory without the caption.
  3. Table 1: report exact n and whether the paired t is two-sided; a one-line note on multiple-comparison handling across the ladder would help.
  4. §6.1: the hinge depth band (30–85%) and λ grid are free parameters; a one-sentence ablation or default-choice justification would reduce the free-parameter list.
  5. Reproducibility Statement: the public repository link is promised for the arXiv update; ensure the three-tier package (no-GPU / 24GB / 96GB) is actually present before camera-ready.
  6. Typos / style: “isassembledinside” (abstract), “genuinestructural”, “difference-in-differences (DiD) trajectories” is fine but DiD is sometimes written without expansion on first use in §3.

Circularity Check

0 steps flagged

No significant circularity: claims are empirical measurements with explicit de-circularization (intact-model dips, causal patchscopes, out-of-domain lens), not results forced by definition or self-citation.

full rationale

This is an empirical/mechanistic paper, not a first-principles derivation. The wrong-dip is defined from layer-wise DiD trajectories (margin_a − margin_b, referenced to the model’s own final DiD sign) and then independently stress-tested: (i) mid-layer states transplanted into neutral contexts drive the wrong token (patchscopes, n=278, 12 ladder models); (ii) dips used to predict compression flips are always measured on the intact model before damage; (iii) late residual attenuation selectively flips high-dip items; (iv) training effects optimized on logit-lens margins are re-measured with causal patchscopes, cutting the causal dip 67–70% while output-only SFT worsens it; (v) an out-of-domain tuned lens restores the dip that an in-domain lens shrinks. These steps break the main circularity risks (lens circularity, post-damage circularity, probe learning the phenomenon). There are no load-bearing self-citations by the author, no uniqueness theorems imported from prior own work, no fitted parameters renamed as predictions of the same quantity, and no mere renaming of a known result without new item-level causal/predictive content. The DiD’s reference to final output is a measurement design choice, not a claim that Y is derived from X when X is defined as Y. Scope limits (free-generation stratification left open, retired ladder tier) are external-validity issues, not circularity. Score 0 is therefore the correct finding.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 3 invented entities

The paper is empirical-mechanistic rather than axiomatic. Load-bearing choices are operational: DiD over polarity pairs as the preference measure, patchscope transplantation as the causal instrument, fixed depth windows for dip statistics, and structural damage operators as probes of late rescue. No new physical entities are postulated; ‘wrong-dip’ and ‘late rescue’ are names for measured trajectories and an account of them. Free parameters are mostly analysis/training hyperparameters that define the instruments, not fitted constants that force the main claim.

free parameters (4)
  • max_wrong_dip depth window (25–90% relative depth)
    Author-chosen window for the deepest wrong-direction DiD excursion; the phenomenon claim depends on this operational definition of the dip statistic.
  • dip-regularization hinge depth band (30–85%) and λ ∈ {0.05, 0.2, 0.5}
    Training penalty location and strength are chosen hyperparameters; dose-response is reported, but the intervention design is not uniquely determined by theory.
  • late residual attenuation fraction (last 15% of blocks) and α grid
    Causal deprivation experiment depends on which late fraction is scaled; results are ordered as predicted but the cut is a free design choice.
  • compression ranks / prune fractions / block-drop fraction (e.g., SVD r64/r128, 50% channel prune, 12.5% depth drop)
    Damage intensities are selected to produce measurable flips; double dissociation is robust across operators but exact ranks are free parameters of the audit.
axioms (5)
  • domain assumption Layer-wise DiD of polarity-controlled minimal pairs cancels item-level lexical frequency and phrasing confounds so residual sign tracks internal preference relative to the model’s own final decision.
    Stated in §3.2 as the justification for DiD over raw margins; central to all trajectory claims.
  • domain assumption Transplanting a mid-layer hidden state into a neutral decoding context (patchscopes) reveals the token preference causally carried by that representation.
    §3.3 and all cross-family causal claims rest on this activation-transplantation semantics.
  • domain assumption Late-layer low-rank compression, block dropping, and structured pruning preferentially consume a late correction budget, whereas whole-depth quantization damages more diffusely.
    Mechanistic interpretation of the double dissociation in §5.2–5.3 and §5.5; used to predict which flips should be dip-marked.
  • domain assumption Single-token or semantic-candidate logprob readouts, when length-controlled at the DiD level, are valid bridges from internal preference to naturalistic vignette decisions.
    §5.6 Acts 2–3; free-generation is treated as a separate interface layer under this assumption.
  • standard math Standard transformer residual-stream decoding through final norm and unembedding yields meaningful intermediate margins (logit/tuned lens).
    Background interpretability practice cited via logit lens and tuned lens; used throughout trajectory measurement.
invented entities (3)
  • wrong-dip (max_wrong_dip and commit_layer_frac) independent evidence
    purpose: Item-level statistics summarizing mid-depth wrong-direction preference and the last disagreeing depth.
    Named operational metrics introduced by the paper; they are definitions of measured trajectories rather than new latent objects, but the central claims are stated in terms of them.
  • late-rescue production process / late correction budget no independent evidence
    purpose: Mechanistic account that mid-layer wrong preference is corrected by late layers and that structural damage exhausts that correction.
    Interpretive construct unifying dip, deprivation, and structural-flip results; supported by causal attenuation but not independently measured as a separate resource outside these experiments.
  • dip-blind answer-interface layer no independent evidence
    purpose: Separate failure mode for free-form/label binding that is not stratified by the internal dip.
    Introduced in §5.6 to explain why generative readouts fail differently from semantic-candidate decisions; useful decomposition, but interface failure is characterized mainly within this paper’s readout ladder.

pith-pipeline@v1.1.0-grok45 · 17342 in / 4139 out tokens · 36774 ms · 2026-07-11T15:57:43.077417+00:00 · methodology

0 comments
read the original abstract

We study how correctness is assembled inside aligned language models, not only whether the final answer is right. Using layer-wise difference-in-differences (DiD) trajectories over polarity-controlled minimal pairs, we identify the wrong-dip: in mid layers (25-90% depth), internal preference transiently commits to the incorrect answer and is rescued only by late-layer correction. We verify this causally with patchscope-style activation transplantation across 17 models, three families, and 64x scale (0.5B-32B). Four findings follow. (1) Alignment amplification of the causal wrong-dip is recipe-specific and emergent: it emerges at 3B in Qwen2.5, remains high, and peaks at 32B (paired t up to 9.7), reverses in Llama-3-8B (t=-2.31), and sits between for Mistral-7B. (2) The dip predicts real compression failures: high-dip items are 3-7x more likely to flip under late-layer low-rank compression, block dropping, or structured pruning, while quantization flips are dip-blind, a double dissociation confirmed by late-layer ablation. (3) The dip is trainable: a LoRA fine-tune with a mid-layer wrong-margin penalty matches output-only SFT accuracy while cutting the causal dip by 67-70% and improving compression robustness; output-only SFT worsens the causal dip by up to 2.8x at perfect surface accuracy. (4) With controlled readouts, the phenomenon survives natural-language I/O: dip stratification of structural-damage failures is significant on naturalistic vignettes, and free-form fragility separates into a dip-auditable late-rescue layer and a dip-blind interface layer. Together, output-level correctness can hide a late-rescue process that governs compression risk, post-training quality, and evaluation distortion.

Figures

Figures reproduced from arXiv: 2607.04640 by Jiaqi Deng.

Figure 1
Figure 1. Figure 1: Overview. Mid layers transiently prefer the wrong answer (the wrong-dip, audited by two item-level statistics); late layers rescue the decision; a readout interface then binds the internal decision to output. Structural damage removes the rescue and flips high-dip items selectively (§5); the generative interface fails in a separate, dip-blind way (§5.6). Phenomenon and metric (§3). On polarity-controlled m… view at source ↗
Figure 2
Figure 2. Figure 2: The two item-level statistics on a schematic DiD trajectory: max_wrong_dip (deepest wrong-direction excursion in the 25–90% depth window) and commit_layer_frac (last depth whose sign disagrees with the output). The final decision is assembled by the late rescue. kens (𝑡pos, 𝑡neg), which appear symmetrically across contexts (surface control). A safety-relevant subset (harm/law/deception/security/trust) is f… view at source ↗
Figure 3
Figure 3. Figure 3: 3.3 Robustness: lens artifacts and causal verification Three controls address measurement validity. (i) An in-domain tuned lens shrinks the dip—but a tuned lens trained on out-of-domain general text restores it, indicating the in-domain lens over-corrects by learning the phenomenon itself. (ii) Causal patchscopes: transplanting the mid-layer hidden state into a neutral decoding context shows the state caus… view at source ↗
Figure 3
Figure 3. Figure 3: Measured layer-wise DiD trajectories on negation/value-flip minimal pairs. Mid layers transiently prefer the wrong answer before late layers rescue the final decision [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Causal verification. Mid-layer hidden states transplanted into a neutral context drive the wrong token, confirming the dip is carried by the representation itself rather than a lens artifact. Three observations. First, alignment amplification of internal wrongness is emergent with scale within a recipe: it emerges at 3B (+0.140), remains high at 7B, and reaches its largest values at 14B/32B—emergence and p… view at source ↗
Figure 5
Figure 5. Figure 5: The causal-dip scale ladder. Alignment amplification emerges at 3B in Qwen2.5, remains high, and peaks at 32B; Llama-3-8B reverses; Qwen3-8B trends with Llama. Output accuracy is matched everywhere. 5 What the Dip Predicts—and What It Does Not 5.1 Negative control: the dip does not index surface brittleness Item-level Spearman correlation between dip and prefix-perturbation flip rate is ≈0.08 at 0.5B; the … view at source ↗
Figure 6
Figure 6. Figure 6: Selective failure under structural damage. Across operators (deprivation, late-SVD, mid-pruning), tasks (negation, role binding, semi-open vignettes) and families, items in the high-dip tertile of the intact model flip far more often than low-dip items. Statistics in §5.2–5.6. 3-bit RTN quantization are dip-blind (item-level 𝜌 ≈ 0 in all four models): low-bit damage is diffuse across depth, so no layer-loc… view at source ↗
Figure 7
Figure 7. Figure 7: The bridge ladder (§5.6, Act 3). (a) Clean pair accuracy across the four readout levels on natural vignettes: the semantic-candidate decision (L1) is robust; generative readouts degrade stepwise; free explanation (L4) is floored on base models. (b) Under identical structural damage on Qwen-7B, the generative interface collapses (up to 100% flips) while the L1 semantic decision flips far less—and those flip… view at source ↗
Figure 8
Figure 8. Figure 8: Robustness transfer at 7B: dip-regularized SFT retains 0.943 ± 0.046 accuracy after mid-SVD r64 vs 0.872 ± 0.085 for output-only SFT (240-item strict set, 3 seeds). 6.4 Metric-consistency check: causal re-measurement of all trained arms Because the regularizer optimizes logit-lens margins, lens-measured reductions risk circularity. We re￾measured every trained arm with the causal patchscopes instrument ( … view at source ↗
Figure 9
Figure 9. Figure 9: The dip as a training-time monitor (Qwen2.5-1.5B, five recipes). (a) All arms saturate held-out accuracy by step 30—the output view cannot distinguish recipes. (b) The internal dip keeps separating them, with a dose-response in 𝜆. (c) End-of-training dip forecasts final mid-SVD retention. The signature replicates on Qwen2.5-7B and Llama-3-8B. ( [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Recipe taxonomy. 𝑥: causal alignment amplification (instruct − base); 𝑦: instruct commit depth; point size: RTN-w3 survival; annotation: inversion residual dip. Two anchored clusters (late-patch vs mid-rewrite) with Mistral-7B in between. 7.2 Plasticity duel. Both models learn a forced preference inversion to 100% output accuracy in 150 LoRA steps, but by different mechanisms: Qwen-7B carries residual int… 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

23 extracted references · 10 linked inside Pith

  1. [1]

    Interpreting GPT: the logit lens

    nostalgebraist. Interpreting GPT: the logit lens. Blog post,LessWrong, 2020.https://www.lesswrong.com/ posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens

    2020

  2. [2]

    Eliciting latent predictions from transformers with the tuned lens.arXiv:2303.08112, 2023

    Belrose, N., et al. Eliciting latent predictions from transformers with the tuned lens.arXiv:2303.08112, 2023

    Pith/arXiv arXiv 2023

  3. [3]

    Patchscopes: A unifying framework for inspecting hidden representations of language models.ICML, 2024

    Ghandeharioun, A., et al. Patchscopes: A unifying framework for inspecting hidden representations of language models.ICML, 2024

    2024

  4. [4]

    Locating and editing factual associations in GPT.NeurIPS, 2022

    Meng, K., et al. Locating and editing factual associations in GPT.NeurIPS, 2022

    2022

  5. [5]

    Towards best practices of activation patching in language models: Metrics and methods

    Zhang, F., Nanda, N. Towards best practices of activation patching in language models: Metrics and methods. ICLR, 2024

    2024

  6. [6]

    Overthinking the truth: Understanding how language models process false demonstrations

    Halawi, D., et al. Overthinking the truth: Understanding how language models process false demonstrations. arXiv:2307.09476, 2023

    Pith/arXiv arXiv 2023

  7. [7]

    The hydra effect: Emergent self-repair in language model computations.arXiv:2307.15771, 2023

    McGrath, T., et al. The hydra effect: Emergent self-repair in language model computations.arXiv:2307.15771, 2023

    Pith/arXiv arXiv 2023

  8. [8]

    The unreasonable ineffectiveness of the deeper layers.arXiv:2403.17887, 2024

    Gromov, A., et al. The unreasonable ineffectiveness of the deeper layers.arXiv:2403.17887, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Men,X.,etal.ShortGPT:Layersinlargelanguagemodelsaremoreredundantthanyouexpect.arXiv:2403.03853, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    The truth is in there: Improving reasoning in language models with layer-selective rank reduction.ICLR, 2024

    Sharma, P., Ash, J., Misra, D. The truth is in there: Improving reasoning in language models with layer-selective rank reduction.ICLR, 2024

    2024

  11. [11]

    Emergent abilities of large language models.TMLR, 2022

    Wei, J., et al. Emergent abilities of large language models.TMLR, 2022

    2022

  12. [12]

    Are emergent abilities of large language models a mirage?NeurIPS, 2023

    Schaeffer, R., Miranda, B., Koyejo, S. Are emergent abilities of large language models a mirage?NeurIPS, 2023

    2023

  13. [13]

    LLM.int8(): 8-bit matrix multiplication for transformers at scale.NeurIPS, 2022

    Dettmers, T., et al. LLM.int8(): 8-bit matrix multiplication for transformers at scale.NeurIPS, 2022

    2022

  14. [14]

    GPTQ: Accurate post-training quantization for generative pre-trained transformers.ICLR, 2023

    Frantar, E., et al. GPTQ: Accurate post-training quantization for generative pre-trained transformers.ICLR, 2023

    2023

  15. [15]

    AWQ: Activation-aware weight quantization for LLM compression and acceleration.MLSys, 2024

    Lin, J., et al. AWQ: Activation-aware weight quantization for LLM compression and acceleration.MLSys, 2024

    2024

  16. [16]

    An empirical study of Qwen3 quantization.arXiv:2505.02214, 2025

    Zheng, X., Li, Y., Chu, H., Feng, Y., Ma, X., Luo, J., Guo, J., Qin, H., Magno, M., Liu, X. An empirical study of Qwen3 quantization.arXiv:2505.02214, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    Decoding compressed trust: Scrutinizing the trustworthiness of efficient LLMs under compression

    Hong, J., et al. Decoding compressed trust: Scrutinizing the trustworthiness of efficient LLMs under compression. ICML, 2024. 15

    2024

  18. [18]

    Compressing LLMs: The truth is rarely pure and never simple.ICLR, 2024

    Jaiswal, A., et al. Compressing LLMs: The truth is rarely pure and never simple.ICLR, 2024

    2024

  19. [19]

    Representation engineering: A top-down approach to AI transparency.arXiv:2310.01405, 2023

    Zou, A., et al. Representation engineering: A top-down approach to AI transparency.arXiv:2310.01405, 2023

    Pith/arXiv arXiv 2023

  20. [20]

    LoRA: Low-rank adaptation of large language models.ICLR, 2022

    Hu, E., et al. LoRA: Low-rank adaptation of large language models.ICLR, 2022

    2022

  21. [21]

    Qwen2.5 technical report.arXiv:2412.15115, 2024

    Yang, A., et al. Qwen2.5 technical report.arXiv:2412.15115, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    The Llama 3 herd of models.arXiv:2407.21783, 2024

    Dubey, A., et al. The Llama 3 herd of models.arXiv:2407.21783, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    Mistral 7B.arXiv:2310.06825, 2023

    Jiang, A., et al. Mistral 7B.arXiv:2310.06825, 2023. 16

    Pith/arXiv arXiv 2023