Pith. sign in

REVIEW 4 major objections 4 minor 34 references

Large language models overuse the rhetorical figure of epanorthosis — "not X, but Y" self-correction — because training data and preference tuning reward emphatic phrasing, and the excess can be measured and dialed back toward human rates.

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 07:14 UTC pith:LTCAP7DP

load-bearing objection A transparent, well-written pilot that names a real LLM tic and ships a LoRA dial, but the headline miscalibration numbers rest on a detector that is not valid on human text. the 4 major comments →

arxiv 2607.21498 v2 pith:LTCAP7DP submitted 2026-07-23 cs.CL cs.AI

Artificial Epanorthosis: Why large language models overuse a classical rhetorical figure, and how to mitigate it

Federico Boggia This is my paper
classification cs.CL cs.AI
keywords epanorthosisrhetorical figureslarge language modelsRLHFLoRAcontrollable text generationstylistic calibrationAI-text detection
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.

Epanorthosis — the rhetorical move of negating a word to replace it with a stronger one, as in "This is not a course. It is a journey of transformation" — has become a signature tic of AI-generated writing. The paper argues that this overuse is a trained disposition: the training distribution and preference tuning reward confident, emphatic phrasing, while left-to-right generation merely surfaces the pattern unedited. It introduces a genre-relative Epanorthosis Index (density divided by the human rate) and reports two-sided miscalibration — overshoot in oratory, undershoot in informal Q&A. It then shows the excess is correctable: a one-line instruction cuts the figure by 70–72 percent in Italian, and a lightweight LoRA adapter can nearly eliminate it, with a scaling dial that lands on the human rate. The stakes are that we may begin to write like the machines, so calibrating style to human registers matters beyond aesthetics.

Core claim

The paper argues that when a model writes "This is not a course. It is a journey of transformation," it is reproducing a figure Cicero and Quintilian catalogued, not inventing a tic: epanorthosis, the self-correction that negates a term to replace it with a stronger one. The central claim is that LLMs overuse this figure because it is a trained disposition — training corpora are rich in promotional prose and preference tuning rewards confident, emphatic phrasing — while left-to-right generation is only an amplifier. Using an Epanorthosis Index (model density divided by human density per ten thousand words), the paper measures two-sided miscalibration: models overshoot in oratory (33.5 vs 14.

What carries the argument

The load-bearing machinery is the Epanorthosis Index: for a genre g, EI_g = D_g/H_g, the model's emphatic-epanorthosis density (per ten thousand words, counting the "not X, but Y" family) divided by the human baseline density. An index of 1 is human-like calibration, above 1 overshoot, below 1 undershoot. The paper's theoretical bridge is Fontanier's classification of epanorthosis as a figure of thought, which licenses measuring the figure in a statistical model regardless of intention. On the mitigation side, the central mechanism is a low-rank adaptation (LoRA) adapter trained by supervised fine-tuning on de-emphasised paraphrases; scaling its contribution by a coefficient alpha at inferen

Load-bearing premise

The human baseline rates that define "overuse" rest on the paper's own detector, which it reports as far less reliable on human text than on model text; if those baselines are inflated, the central overshoot finding shrinks or disappears.

What would settle it

Hand-annotate a human oratory sample with the same criteria used for model text and recompute the Epanorthosis Index with the corrected human density; if the corrected human rate is not below the model rate, the paper's signature overshoot result is a detector artifact.

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

If this is right

  • The same training stages meant to make models helpful — instruction tuning and preference optimization — are what bake in the rhetorical excess, so fixing the data and reward signal would address the cause rather than the symptom.
  • A cheap one-line instruction cuts emphatic self-correction by 70–72 percent in Italian oratory and argument, so practical mitigation is available even without retraining.
  • A supervised LoRA adapter nearly removes the figure, and its scaling coefficient can settle output at the human rate rather than at zero — the paper's definition of proper calibration.
  • Measuring style genre-by-genre reveals two-sided errors that a raw frequency count hides, so evaluations of AI writing should use register-relative baselines.
  • If left unchecked, the miscalibrated style can transfer to human writing, which is the paper's stated reason to care.

Where Pith is reading between the lines

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

  • Editorial inference: Because the detector's precision on human text is only 0.17, the human densities in the baseline table are likely inflated by false positives; recomputing the Epanorthosis Index with human-validated labels could reduce or even reverse the oratory overshoot.
  • Editorial inference: The attribution of the effect to RLHF is extrapolated, not directly tested; a base-model-versus-aligned comparison across several developer model families would confirm or refute it.
  • Editorial inference: If epanorthosis is a figure of thought that models rediscover statistically, other named figures (e.g., chiasmus, epiphora) should also show register-dependent, detectable frequencies, giving AI-text detection a whole-rhetoric fingerprint.
  • Editorial inference: The supervised adapter's near-total elimination of the figure at full strength may suppress legitimate correction in genres like argument; content-fidelity and appropriateness evaluation, prescribed but not yet reported, is needed before deployment.

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 / 4 minor

Summary. The paper argues that LLMs systematically overuse epanorthosis ('Not X. Y' self-correction) because of a training distribution rich in promotional prose and because RLHF rewards confident phrasing, with left-to-right generation only an amplifier. It proposes an Epanorthosis Index (density relative to a genre-specific human baseline), reports an English pilot study claiming bidirectional miscalibration (overshoot in oratory, undershoot in informal Q&A), and presents mitigation results: a LoRA adapter with an α-scaling 'dial', a one-line prompting intervention in Italian, and training-free best-of-n/rewrite baselines. The paper is candid about limitations, including per-genre p-values that do not survive multiple-comparison correction, detector precision asymmetry, small samples, and the absence of a base-model-versus-aligned comparison.

Significance. If the measurement were reliable, the paper would make a useful conceptual contribution: it connects a classical rhetorical figure to a measurable stylistic property of LLMs, proposes a register-relative calibration target rather than simple suppression, and provides reproducible artifacts (LoRA recipe, adapter weights, evaluation scripts, a negative DPO result). The framing of mitigation as calibration to human rates rather than elimination is thoughtful. However, the central quantitative claims currently rest on pilot-scale evidence with a detector that is much noisier on human text than on model text, uncorrected multiple testing, and no significance tests for the mitigation tables. The significance is therefore conditional: the paper is a promising research programme, not yet an established finding.

major comments (4)
  1. [§4.1, Eq. (EI), Appendix B, Table B.2] The detector's precision asymmetry is load-bearing for the central EI ratios. Table B.2 reports precision 0.17 on human windows versus 0.82 on model windows. Since EI_g = D_g/H_g divides by the human baseline H_g, an inflated H_g (due to ordinary adversatives misclassified as epanorthosis) directly deflates every EI value in Table 1. This could erase or reverse the reported overshoot/undershoot pattern. The English detector is not separately validated, so the Italian figures are the only evidence. The manuscript acknowledges the asymmetry but does not bound, correct, or sensitivity-test its effect on Table 1. I request a precision/recall-adjusted estimate of H_g, a sensitivity analysis, or a re-annotation of the human windows before the EI claims can be accepted.
  2. [§4.1, Limitations (Statistical power and multiplicity)] The abstract and Section 4.1 claim 'mis-calibration by register in both directions' with oratory p=0.03 and informal Q&A p=0.05. The Limitations state that under strict multiple-comparison correction neither result survives. As reported, the two headline effects are at best suggestive. The abstract and conclusion should be tempered to explicitly say these are exploratory, and Table 1 should report corrected p-values or clearly mark the uncorrected values. This is not a fatal flaw given the paper's transparency, but the central claim of bidirectional miscalibration is not currently established at conventional significance levels.
  3. [§7.2, Table 2; §7.9, Table 5] The mitigation claims are based on point estimates without significance tests: six generations per prompt in Table 2 and per-genre baselines resting on two prompts in Table 5. The 'dial' claim—specifically that argument reaches the human rate at α=0.75—rests on a single point estimate (12.3) with no confidence interval. The promotional spike at α=0.25 (96.0, above the α=0 baseline of 61.4) could be noise or a real non-monotonic effect; without uncertainty bounds it cannot be interpreted. Additionally, the content-fidelity evaluation prescribed in §7.7 is not reported, so the adapter's principal risk remains unmeasured. Please provide confidence intervals, more generations, or a significance test, and report the content-fidelity results before presenting the adapter as a validated calibration tool.
  4. [§3, Limitations (Causal attribution, Single model family)] The abstract states that overuse is 'a trained disposition, driven mainly by training distribution and preference tuning (RLHF)', but the paper's own measurements do not test this causal claim: there is no base-model-versus-aligned comparison, and all three models come from one alignment pipeline. The section correctly labels the RLHF attribution as inferential, but this caveat is absent from the abstract and conclusion. Either soften the causal attribution to an explicitly stated hypothesis, or add the base/aligned comparison that would test it. This is load-bearing because the paper's framing ('trained disposition') is presented as an explanation, not merely a conjecture.
minor comments (4)
  1. [§4.1, 'specificity check'] The claim that models use fewer neutral connectives (1.5 vs. 7.5 in oratory) supports the interpretation that the spike is not marker-heaviness, but no significance test or window-count is reported. Please provide the underlying comparison or mark it as descriptive.
  2. [§7.2, 'T wo pilots'] There is a stray space in 'T wo pilots'. Throughout the manuscript there are similar ligature/formatting artifacts (e.g., 'difficult', 'efficiently', 'sufficient', 'T able'). These should be cleaned in the production version.
  3. [§7.2 vs. §4.1] The model mismatch is acknowledged, but Section 7.2's adapter is trained on Qwen2.5-7B-Instruct while the measurement and prompting demonstrations use Claude Haiku/Sonnet/Opus. This limits the direct comparability of the α-dial calibration in Table 2 to the human baselines in Table 1. Please state this explicitly in the main text rather than only in Limitations.
  4. [Appendix B, Table B.2] The per-genre precision/recall rows for Speech, Narrative, and Social rest on very few true instances (seven or fewer). This is acknowledged in the text, but the table itself would benefit from an explicit 'n' column showing the number of true instances per row.

Circularity Check

0 steps flagged

No circular reasoning: the Epanorthosis Index is a definitional ratio used on measured densities, and the mitigation results are empirical demonstrations rather than predictions forced by the same inputs.

full rationale

The paper's quantitative core is a set of empirical density comparisons, not a derivation. EI_g = D_g/H_g is explicitly defined as 'density relative to the human rate' (Section 4.1), so the calibration claims are readings of a metric, not consequences of an equation that assumes them. The LoRA dial is a swept control variable: Table 2 reports densities across α values, and the observation that argument prose reaches 12.3 at α=0.75 is an empirical calibration point, not a fitted parameter later renamed as a prediction. The 'one-line instruction' and rewrite results are manipulation checks measured with the same operationalization used to define the figure; this is consistent metric use, not circularity. The best-of-n method transparently selects the candidate with the lowest density, so its reported reduction is a selection rule rather than a hidden prediction; the paper does not present it as an independent discovery. There are no load-bearing self-citations: the rhetorical-divergence support [1] is an external PNAS study, and the RLHF hypothesis rests on the independent Juzek-Ward lexical study with the paper's own caveat that the attribution is inferential. The acknowledged detector asymmetry (precision 0.17 on human windows vs 0.82 on model windows, Appendix B) and the uncorrected multiplicity of the Section 4.1 tests are genuine validity threats to the human baselines and p-values, but they are measurement-quality limitations, not reductions of a derivation to its own inputs. The limitations section itself flags these issues, so the paper is not hiding the weakness. No step satisfies the requirement of exhibiting an equation that is identical to its input by construction or a fitted parameter renamed as a prediction.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central claims rest on a hand-chosen detector, arbitrary windowing, and alpha dial settings tuned to human baselines. No new physical or theoretical entities are introduced; the Epanorthosis Index is a metric, not an entity.

free parameters (3)
  • Detector trigger set for emphatic-correction frames
    The detector counts only the hand-selected frames ('not X, but Y', 'not only X but Y', sentence-split 'Not X. Y'). This is a modeling choice; it excludes other epanorthosis markers (e.g., 'cioè', 'diciamo'), so all density and EI values depend on this arbitrary trigger set.
  • LoRA scaling coefficient α per genre = α=0.75 for argument (target human rate 12); α≈0.4 for oratory (inferred)
    The dial setting that lands on the human rate is chosen after seeing the data. It is a fitted knob, not a predicted parameter, and the promotional register even shows non-monotonic behavior at α=0.25.
  • Window size for density counts = few hundred words (not specified exactly)
    Window length is chosen arbitrarily and affects per-window density and the permutation-test results.
axioms (5)
  • domain assumption Epanorthosis can be identified and measured without access to the speaker's intention (Fontanier's figure-of-thought classification).
    Section 2 uses this classification to transfer the concept to LLM output; the status of intention is theoretically contested.
  • domain assumption The surface construction 'Not X. Y' is an adequate proxy for epanorthosis for measurement.
    Section 4 explicitly acknowledges it measures a proxy, not the figure of thought, but proceeds to draw quantitative conclusions from that proxy.
  • domain assumption Human historical baselines (19th- and early 20th-century texts) represent the relevant human rate for modern genre comparison.
    Section 4.1 caveats note era mismatch for oratory, argument, and narrative; the overshoot in oratory could partly reflect period rhetoric.
  • domain assumption Observations on one model family (Claude Haiku/Sonnet/Opus) can inform claims about instruction-tuned models generally.
    The Limitations state the three sizes share a single alignment pipeline, so effects cannot be separated from that pipeline's idiosyncrasies.
  • standard math Permutation tests are a valid significance test for per-window density differences.
    Section 4.1 uses 5,000-iteration label-permutation tests; this assumes exchangeability under the null and independent windows.

pith-pipeline@v1.3.0-alltime-deepseek · 16611 in / 12038 out tokens · 114346 ms · 2026-08-01T07:14:39.621668+00:00 · methodology

0 comments
read the original abstract

A rhetorical figure that Cicero and Quintilian catalogued two thousand years ago reappears, systematically, in the text of large language models: epanorthosis, the self-correction of the specimen {\guillemotleft}This is not a course. It is a journey of transformation{\guillemotright}. This essay argues that the overuse is a trained disposition, driven mainly by a training distribution rich in promotional prose and by preference tuning (RLHF) that rewards confident, emphatic phrasing; the left-to-right nature of generation is an amplifier rather than the root cause. Building on evidence that models diverge from human rhetorical style, and on Fontanier's classification of epanorthosis as a figure of thought, it sets out a programme that scores the figure against genre-specific human baselines through an Epanorthosis Index (density relative to the human rate). A first measurement, on three sizes of one instruction-tuned model family, finds mis-calibration by register in both directions: the models overshoot in oratory (about twofold, near threefold in Italian, concentrated in the larger tiers) and undershoot in informal question-and-answer writing, while matching humans in argument, journalism, and encyclopedic prose. Three constructive contributions follow: a survey of mitigation techniques centred on lightweight LoRA adapters; a demonstration, in Italian, that a one-line instruction cuts the figure by half to nearly three-quarters and that a supervised-fine-tuning adapter removes it almost entirely, with a scaling coefficient that dials the reduction back onto the human rate; and the argument that the target is calibration to the human rate for each genre, not elimination. It closes on the stakes: the real risk is that we begin to write like the machines.

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

34 extracted references · 26 linked inside Pith

  1. [1]

    Reinhart, A., Markey, B., Laudenbach, M., Pantusen, K., Yurko, R., Weinberg, G., & Brown, D. W. (2025). Do LLMs write like humans? Variation in grammatical and rhetorical styles. Proceedings of the National Academy of Sciences , 122(8), e2422455122. arXiv:2410.16107

    Pith/arXiv arXiv 2025

  2. [2]

    Fontanier, P. (1827). Les Figures du discours (ed. G. Genette, Paris: Flammarion, 1968)

    1968

  3. [3]

    Institutio Oratoria 9.1.30; 9.2.17–18, 59–60; 9.3.88–89 (c

    Quintilian. Institutio Oratoria 9.1.30; 9.2.17–18, 59–60; 9.3.88–89 (c. 95 CE) (Loeb Classical Library, Harvard University Press)

  4. [4]

    In Catilinam I.2 and II.1 (63 BCE) (Loeb Classical Library, Harvard University Press)

    Cicero. In Catilinam I.2 and II.1 (63 BCE) (Loeb Classical Library, Harvard University Press)

  5. [5]

    Inferno, Canto XXXIII, l

    Dante Alighieri. Inferno, Canto XXXIII, l. 75

  6. [6]

    Lausberg, H. (1998). Handbook of Literary Rhetoric: A Foundation for Literary Study . Leiden: Brill

    1998

  7. [7]

    N., Kaiser, Ł., & Polo- sukhin, I

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polo- sukhin, I. (2017). Attention Is All You Need. NeurIPS 2017 . arXiv:1706.03762

    Pith/arXiv arXiv 2017

  8. [8]

    Holtzman, A., Buys, J., Du, L., Forbes, M., & Choi, Y. (2020). The Curious Case of Neural Text Degeneration. ICLR 2020 . arXiv:1904.09751

    Pith/arXiv arXiv 2020

  9. [9]

    Ouyang, L., Wu, J., Jiang, X., et al. (2022). Training language models to follow instructions with human feedback. NeurIPS 2022 . arXiv:2203.02155

    Pith/arXiv arXiv 2022

  10. [10]

    S., & Ward, Z

    Juzek, T. S., & Ward, Z. B. (2025). Why Does ChatGPT “Delve” So Much? Exploring the Sources of Lexical Overrepresentation in Large Language Models. Proceedings of COLING 2025, 6397–6411. arXiv:2412.11385

    Pith/arXiv arXiv 2025

  11. [11]

    Kobak, D., González-Márquez, R., Horvát, E.-Á., & Lause, J. (2025). Delving into LLM-assisted writing in biomedical publications through excess vocabulary. Science Advances, 11(27). Preprint arXiv:2406.07016 (2024)

    Pith/arXiv arXiv 2025

  12. [12]

    Liang, W., Izzo, Z., Zhang, Y., et al. (2024). Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT on AI Conference Peer Reviews. ICML 2024 . arXiv:2403.07183

    Pith/arXiv arXiv 2024

  13. [13]

    Gehrmann, S., Strobelt, H., & Rush, A. M. (2019). GLTR: Statistical Detection and Visualization of Generated Text. ACL 2019, System Demonstrations . arXiv:1906.04043

    Pith/arXiv arXiv 2019

  14. [14]

    D., & Finn, C

    Mitchell, E., Lee, Y., Khazatsky, A., Manning, C. D., & Finn, C. (2023). DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature. ICML 2023 . arXiv:2301.11305

    Pith/arXiv arXiv 2023

  15. [15]

    Guo, B., Zhang, X., Wang, Z., et al. (2023). How Close is ChatGPT to Human Experts? Comparison Corpus, Evaluation, and Detection. arXiv:2301.07597

    Pith/arXiv arXiv 2023

  16. [16]

    Mauri, C., Ballarè, S., Goria, E., Cerruti, M., & Suriano, F. (2019). KIParla corpus: a new resource for spoken Italian. Proceedings of the 6th Italian Conference on Computational Linguistics (CLiC-it 2019)

    2019

  17. [17]

    Dubremetz, M., & Nivre, J. (2018). Rhetorical Figure Detection: Chiasmus, Epanaphora, Epiphora. Frontiers in Digital Humanities , 5:10

    2018

  18. [18]

    Bothwell, S., DeBenedetto, J., Crnkovich, T., Müller, H., & Chiang, D. (2023). Introducing Rhetor- ical Parallelism Detection: A New Task with Datasets, Metrics, and Baselines. EMNLP 2023 . arXiv:2312.00100. 14

    Pith/arXiv arXiv 2023

  19. [19]

    Wei, J., Wang, X., Schuurmans, D., et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. NeurIPS 2022 . arXiv:2201.11903

    Pith/arXiv arXiv 2022

  20. [20]

    S., Reid, M., Matsuo, Y., & Iwasawa, Y

    Kojima, T., Gu, S. S., Reid, M., Matsuo, Y., & Iwasawa, Y. (2022). Large Language Models are Zero-Shot Reasoners. NeurIPS 2022 . arXiv:2205.11916

    Pith/arXiv arXiv 2022

  21. [21]

    Madaan, A., Tandon, N., Gupta, P., et al. (2023). Self-Refine: Iterative Refinement with Self- Feedback. NeurIPS 2023 . arXiv:2303.17651

    Pith/arXiv arXiv 2023

  22. [22]

    Huang, J., Chen, X., Mishra, S., et al. (2024). Large Language Models Cannot Self-Correct Reason- ing Yet. ICLR 2024 . arXiv:2310.01798

    Pith/arXiv arXiv 2024

  23. [23]

    Bai, Y., Kadavath, S., Kundu, S., et al. (2022). Constitutional AI: Harmlessness from AI Feedback. arXiv:2212.08073

    Pith/arXiv arXiv 2022

  24. [24]

    Searle, J. R. (1980). Minds, Brains, and Programs. Behavioral and Brain Sciences , 3(3), 417–457

    1980

  25. [25]

    J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., & Chen, W

    Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., & Chen, W. (2022). LoRA: Low-Rank Adaptation of Large Language Models. ICLR 2022 . arXiv:2106.09685

    Pith/arXiv arXiv 2022

  26. [26]

    Dettmers, T., Pagnoni, A., Holtzman, A., & Zettlemoyer, L. (2023). QLoRA: Efficient Finetuning of Quantized LLMs. NeurIPS 2023 . arXiv:2305.14314

    Pith/arXiv arXiv 2023

  27. [27]

    D., & Finn, C

    Rafailov, R., Sharma, A., Mitchell, E., Ermon, S., Manning, C. D., & Finn, C. (2023). Di- rect Preference Optimization: Your Language Model is Secretly a Reward Model. NeurIPS 2023 . arXiv:2305.18290

    Pith/arXiv arXiv 2023

  28. [28]

    Dathathri, S., Madotto, A., Lan, J., et al. (2020). Plug and Play Language Models: A Simple Approach to Controlled Text Generation. ICLR 2020 . arXiv:1912.02164

    Pith/arXiv arXiv 2020

  29. [29]

    D., McCann, B., et al

    Krause, B., Gotmare, A. D., McCann, B., et al. (2021). GeDi: Generative Discriminator Guided Sequence Generation. Findings of EMNLP 2021 . arXiv:2009.06367

    Pith/arXiv arXiv 2021

  30. [30]

    Yang, K., & Klein, D. (2021). FUDGE: Controlled Text Generation With Future Discriminators. NAACL 2021 . arXiv:2104.05218

    Pith/arXiv arXiv 2021

  31. [31]

    M., Thiergart, L., Leech, G., et al

    Turner, A. M., Thiergart, L., Leech, G., et al. (2023). Steering Language Models With Activation Engineering (the ActAdd method). arXiv:2308.10248

    Pith/arXiv arXiv 2023

  32. [32]

    Zou, A., Phan, L., Chen, S., et al. (2023). Representation Engineering: A Top-Down Approach to AI Transparency. arXiv:2310.01405

    Pith/arXiv arXiv 2023

  33. [33]

    Li, K., Patel, O., Viégas, F., Pfister, H., & Wattenberg, M. (2023). Inference-Time Intervention: Eliciting Truthful Answers from a Language Model. NeurIPS 2023 . arXiv:2306.03341

    Pith/arXiv arXiv 2023

  34. [34]

    Not X. Y

    Sharma, M., Tong, M., Korbak, T., et al. (2023). Towards Understanding Sycophancy in Language Models. arXiv:2310.13548. Appendix A. A reproducible mitigation recipe Two training scripts accompany Section 7.2. The first attempt ( artificial-epanorthosis-lora- recipe.py) reads the delivered preference pairs, built separately by the counterfactual augmen- ta...

    Pith/arXiv arXiv 2023