arxiv: 2603.13683 · v3 · submitted 2026-03-14 · 💻 cs.CL · cs.AI· cs.CY

Recognition: 2 theorem links

· Lean Theorem

Preconditioned Test-Time Adaptation for Out-of-Distribution Debiasing in Narrative Generation

Hanwen Shen , Ting Ying , Jiajie Lu , Shanshan Wang

Authors on Pith no claims yet

Pith reviewed 2026-05-15 12:10 UTC · model grok-4.3

classification 💻 cs.CL cs.AIcs.CY

keywords test-time adaptationdebiasingout-of-distributionnarrative generationLoRApreconditionerbias detection

0 comments

The pith

CAP-TTA applies preconditioned test-time adaptation to reduce bias in LLM narrative generation on unfamiliar prompts.

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

The paper shows that debiased LLMs still fail on high-bias out-of-distribution prompts because those prompts shift the input distribution. CAP-TTA detects this shift with a bias-risk score and applies targeted LoRA updates only when needed, using a precomputed diagonal preconditioner for fast, stable optimization. This approach cuts toxicity and bias scores while keeping narrative fluency high and avoiding the forgetting that comes with full retraining. A sympathetic reader would care because it offers a practical way to make language models safer in real-world use without heavy computation at every step.

Core claim

CAP-TTA triggers context-aware LoRA updates only when a bias-risk score exceeds a threshold, and uses an offline precomputed diagonal preconditioner to ensure fast and stable optimization during test-time adaptation for debiasing narrative generation on OOD prompts.

What carries the argument

CAP-TTA framework: context-aware LoRA updates triggered by bias-risk score threshold, accelerated by offline precomputed diagonal preconditioner for stable test-time optimization.

If this is right

Reduces toxicity and bias scores on high-bias OOD prompts compared to static models.
Achieves lower latency than standard optimizers like AdamW or SGD.
Prevents catastrophic forgetting during adaptation.
Improves narrative fluency over baselines without losing debiasing effectiveness.

Where Pith is reading between the lines

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

Such selective adaptation could extend to other safety concerns like factual accuracy in generated stories.
Precomputing the preconditioner offline might allow deployment on resource-limited devices for real-time correction.
Threshold tuning on the bias-risk score could be generalized to other distribution shift detectors in language tasks.

Load-bearing premise

The assumption that an offline precomputed diagonal preconditioner combined with a bias-risk score threshold will reliably detect and correct distribution shifts on high-bias OOD prompts without introducing instabilities.

What would settle it

A set of OOD prompts where applying CAP-TTA either fails to lower the bias score below baseline levels or increases latency beyond standard methods.

Figures

Figures reproduced from arXiv: 2603.13683 by Hanwen Shen, Jiajie Lu, Shanshan Wang, Ting Ying.

**Figure 1.** Figure 1: Static generation vs. prior test-time adaptation (TTA) vs. CAP-TTA. Static generation uses frozen parameters. Prior TTA performs online updates during generation, which can be costly and unstable. CAP-TTA decouples adaptation into an offline precomputed preconditioner P0 and an online bias-triggered, lightweight preconditioned update (optionally routed to a safe corpus with 4 types) when the trigger score … view at source ↗

**Figure 2.** Figure 2: This is evaluated using the bias trigger score. Per-prompt bias trajectories over narrative segments on the [PITH_FULL_IMAGE:figures/full_fig_p020_2.png] view at source ↗

**Figure 3.** Figure 3: This is evaluated using the bias trigger score. This figure shows triggering trade-off for CAP-TTA. We [PITH_FULL_IMAGE:figures/full_fig_p021_3.png] view at source ↗

**Figure 4.** Figure 4: This is evaluated using the bias trigger score. This figure shows comparison of static baselines on toxic [PITH_FULL_IMAGE:figures/full_fig_p022_4.png] view at source ↗

**Figure 5.** Figure 5: This is evaluated using the bias final. This figure shows a empirical CDF of the bias metric [PITH_FULL_IMAGE:figures/full_fig_p023_5.png] view at source ↗

**Figure 6.** Figure 6: This is graded by bias final. This figure shows the distribution of bias/toxicity scores across methods on the [PITH_FULL_IMAGE:figures/full_fig_p024_6.png] view at source ↗

**Figure 7.** Figure 7: Ablation on test-time update latency (ECDF). This figure shows the empirical CDF of the total parameter-update time per prompt for different inference-time strategies on Qwen3. We compare unpreconditioned TTA-SGD (blue) against preconditioned CAP-TTA variants with different trigger thresholds ϵ. Curves further left indicate lower update-time overhead (better efficiency); the plot shows that preconditioning… view at source ↗

**Figure 8.** Figure 8: LoRA Structure. We only update a fraction of weights during TTA. [PITH_FULL_IMAGE:figures/full_fig_p028_8.png] view at source ↗

**Figure 9.** Figure 9: Example of question given to human annotators. [PITH_FULL_IMAGE:figures/full_fig_p031_9.png] view at source ↗

read the original abstract

Although debiased large language models (LLMs) excel at handling known or low-bias prompts, they often fail on unfamiliar and high-bias prompts. We demonstrate via out-of-distribution (OOD) detection that these high-bias prompts cause a distribution shift, degrading static model performance. To enable real-time correction, we propose CAP-TTA, a test-time adaptation framework. CAP-TTA triggers context-aware LoRA updates only when a bias-risk score exceeds a set threshold. By utilizing an offline precomputed diagonal preconditioner, it ensures fast and stable optimization. Across multiple benchmarks and human evaluations, CAP-TTA effectively reduces toxicity/bias score with significantly lower latency than standard optimization methods (e.g., AdamW or SGD). Furthermore, it prevents catastrophic forgetting, and substantially improves narrative fluency over state-of-the-art baselines without compromising debiasing performance.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

CAP-TTA gives a practical triggered LoRA setup for OOD debiasing in narratives using an offline preconditioner, but the performance claims sit on high-level assertions without numbers or shift validation.

read the letter

The main thing here is that the paper introduces CAP-TTA as a test-time adaptation method that only runs context-aware LoRA updates when a bias-risk score crosses a threshold, relying on an offline diagonal preconditioner to keep those updates fast and stable for narrative generation under distribution shift. It targets the gap where static debiased models still produce toxicity on unfamiliar high-bias prompts. The approach avoids full retraining and claims to cut bias scores while preserving fluency and avoiding forgetting, which matches real deployment needs for low-latency fixes. The combination of trigger-based adaptation and preconditioning is a reasonable engineering synthesis that builds on existing TTA and LoRA ideas without obvious circularity in the parameter fitting. The human evaluations and benchmark comparisons are the parts that could make this useful to practitioners if the details hold up. The soft spots are mostly around missing substance. The abstract gives no quantitative results, no error bars, and no derivation for the bias-risk score or how the preconditioner diagonal is actually computed from ID data. That leaves the lower-latency and stability claims hard to judge. The stress-test concern about the fixed preconditioner mismatching curvature on OOD prompts is worth checking directly in the full experiments; if the paper only validates on ID data and then applies it to triggered cases without showing convergence behavior or instabilities, the no-retuning and fast-stable claims rest on untested ground. The tunable threshold and LoRA factors also look like they could need per-scenario adjustment. This is for readers working on safe LLM deployment for creative text who want concrete test-time tricks rather than theory. Someone already familiar with LoRA and OOD detection would get actionable ideas here, though they'd need the full method and results sections to try it. It deserves peer review to surface the implementation details and test whether the preconditioner actually delivers under the shifts it claims to handle.

Referee Report

2 major / 1 minor

Summary. The paper introduces CAP-TTA, a test-time adaptation framework for debiasing LLMs on out-of-distribution high-bias prompts in narrative generation. It detects distribution shifts via a bias-risk score that triggers context-aware LoRA updates only above a threshold, using an offline precomputed diagonal preconditioner to enable fast, stable optimization. The central claims are that this reduces toxicity/bias scores with significantly lower latency than AdamW or SGD, prevents catastrophic forgetting, and improves narrative fluency over SOTA baselines across benchmarks and human evaluations without compromising debiasing.

Significance. If the performance and stability claims hold with supporting data, the work would offer a practical advance in efficient, real-time debiasing for LLMs by avoiding full retraining or per-prompt retuning. The preconditioned TTA approach could be significant for applications requiring safe narrative generation under varying prompt distributions, potentially reducing computational overhead while maintaining fluency.

major comments (2)

[Abstract] Abstract: The central performance claims (reduced toxicity/bias scores, significantly lower latency than AdamW/SGD, improved fluency, no catastrophic forgetting) are stated at a high level without any quantitative results, error bars, benchmark scores, or statistical details, making it impossible to assess effect sizes or verify the claims against baselines.
[Method] Method description: The offline precomputed diagonal preconditioner is presented as ensuring fast and stable LoRA updates on OOD prompts, but no equation, derivation (e.g., as Hessian diagonal approximation from ID data), or validation experiments on high-bias OOD regimes are provided. This is load-bearing for the stability and low-latency assertions, as curvature mismatch under distribution shift could produce ill-conditioned steps or instabilities.

minor comments (1)

[Abstract] The abstract would be strengthened by including at least one key quantitative result (e.g., latency reduction or bias-score delta) to allow readers to gauge the magnitude of the reported improvements.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and will revise the manuscript to improve clarity and completeness.

read point-by-point responses

Referee: [Abstract] Abstract: The central performance claims (reduced toxicity/bias scores, significantly lower latency than AdamW/SGD, improved fluency, no catastrophic forgetting) are stated at a high level without any quantitative results, error bars, benchmark scores, or statistical details, making it impossible to assess effect sizes or verify the claims against baselines.

Authors: We agree that the abstract would benefit from quantitative highlights to better convey effect sizes. In the revision, we will add specific results such as toxicity score reductions (e.g., 25-40% relative improvement), latency comparisons (e.g., 3-5x faster than AdamW), fluency gains, and references to error bars or statistical tests from the experiments, while staying within abstract length limits. revision: yes
Referee: [Method] Method description: The offline precomputed diagonal preconditioner is presented as ensuring fast and stable LoRA updates on OOD prompts, but no equation, derivation (e.g., as Hessian diagonal approximation from ID data), or validation experiments on high-bias OOD regimes are provided. This is load-bearing for the stability and low-latency assertions, as curvature mismatch under distribution shift could produce ill-conditioned steps or instabilities.

Authors: We acknowledge the need for explicit details on the preconditioner. The revision will include the equation (diagonal of the empirical Hessian or Fisher information matrix computed offline on ID data), a brief derivation showing its role in approximating curvature for stable gradient steps, and new validation experiments on high-bias OOD regimes including condition number analysis, convergence curves, and stability metrics to demonstrate robustness against potential curvature mismatch. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation or claims

full rationale

The paper presents CAP-TTA as an empirical method that precomputes a diagonal preconditioner offline from ID data and uses a tunable bias-risk threshold to trigger LoRA updates on OOD prompts. Performance claims (reduced toxicity, lower latency, no catastrophic forgetting, improved fluency) are supported by benchmark results and human evaluations rather than any derivation that reduces predictions or results to quantities defined inside the same equations or fitted parameters. No self-definitional loops, fitted-input predictions, or load-bearing self-citations appear in the provided description; the preconditioner and threshold are external to the adaptation step itself, making the central claims independently falsifiable on external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The framework rests on the assumption that a precomputed diagonal preconditioner stabilizes LoRA updates on distribution-shifted prompts and that a bias-risk score can be computed reliably from context without additional labeled data. No new physical entities are introduced.

free parameters (2)

bias-risk threshold
Value chosen to decide when to trigger LoRA updates; its specific setting is not derived from first principles in the abstract.
LoRA rank and scaling factors
Standard hyperparameters for the adaptation module whose values affect update magnitude.

axioms (2)

domain assumption Test-time LoRA updates on detected OOD prompts can correct bias without catastrophic forgetting of prior capabilities.
Invoked when claiming preservation of narrative fluency alongside debiasing.
domain assumption A diagonal preconditioner computed offline remains effective for online adaptation across multiple high-bias prompts.
Central to the claim of fast and stable optimization versus AdamW or SGD.

pith-pipeline@v0.9.0 · 5451 in / 1558 out tokens · 37210 ms · 2026-05-15T12:10:15.671843+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

By utilizing an offline precomputed diagonal preconditioner, it ensures fast and stable optimization... ϕt+1=ϕt−αt·P0gt
IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean absolute_floor_iff_bare_distinguishability unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We define OOD-ness by treating WritingPrompts as the in-distribution reference... kNN and Mahalanobis indicate RTP is Far-OOD

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

14 extracted references · 14 canonical work pages · 3 internal anchors

[1]

Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski, and Rosanne Liu

Evaluation of text generation: A survey.arXiv preprint arXiv:2006.14799. Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski, and Rosanne Liu. 2020. Plug and play language models: A simple approach to controlled text generation. In International Conference on Learning Representa- tions. Tim Dettmers, Artidoro...

work page arXiv 2006
[2]

Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei A Efros

Bias and fairness in large language models: A survey.Computational Linguistics, 50(3):1097– 1179. Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei A Efros. 2022. Test-time training with masked autoen- coders.Advances in Neural Information Processing Systems, 35:29374–29385. Iker García-Ferrero, David Montero, and Román Orús

work page 2022
[3]

Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A

Refusal steering: Fine-grained control over llm refusal behaviour for sensitive topics.Preprint, arXiv:2512.16602. Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A. Smith. 2020. Realtoxic- ityprompts: Evaluating neural toxic degeneration in language models. InFindings of the Association for Computational Linguistics: EMNLP 2020, pages...

work page arXiv 2020
[4]

strong objectivity?

Depth-wise activation steering for honest lan- guage models.Preprint, arXiv:2512.07667. Kevin A. Hallgren. 2012. Computing inter-rater reliabil- ity for observational data: An overview and tutorial. Tutorials in Quantitative Methods for Psychology, 8(1):23–34. Sandra Harding. 1992. Rethinking standpoint episte- mology: What is “strong objectivity?”.The Ce...

work page arXiv 2012
[5]

Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin

Overcoming catastrophic forgetting in neural networks.Proceedings of the National Academy of Sciences, 114(13):3521–3526. Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin

work page
[6]

Holistic Evaluation of Language Models

A simple unified framework for detecting out- of-distribution samples and adversarial attacks. In Advances in Neural Information Processing Systems. Percy Liang, Rishi Bommasani, Tony Lee, Dimitris Tsipras, Dilara Soylu, Michihiro Yasunaga, Yian Zhang, Deepak Narayanan, Yuhuai Wu, Ananya Ku- mar, Benjamin Newman, Binhang Yuan, Bobby Yan, Ce Zhang, Christi...

work page internal anchor Pith review Pith/arXiv arXiv 2022
[7]

InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), pages 1878–1893

An empirical survey of the effectiveness of debiasing techniques for pre-trained language models. InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), pages 1878–1893. Shira Mitchell, Eric Potash, Solon Barocas, Alexander D’Amour, and Kristian Lum. 2021. Algorithmic fair- ness: Choices, assump...

work page 2021
[8]

InInternational Conference on Learning Representations

Towards stable test-time adaptation in dynamic wild world. InInternational Conference on Learning Representations. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welind...

work page 2022
[9]

Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters

Direct preference optimization: Your language model is secretly a reward model. InAdvances in Neural Information Processing Systems, volume 36. Shauli Ravfogel, Yanai Elazar, Hila Gonen, Michael Twiton, and Yoav Goldberg. 2020. Null it out: Guard- ing protected attributes by iterative nullspace projec- tion. InProceedings of the 58th Annual Meeting of the...

work page internal anchor Pith review Pith/arXiv arXiv 2020
[10]

InProceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 20827–20840

Out-of-distribution detection with deep nearest neighbors. InProceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pages 20827–20840. PMLR. Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei A. Efros, and Moritz Hardt. 2020. Test-time training with self-supervision for generalizat...

work page arXiv 2020
[11]

InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201–7211

Continual test-time domain adaptation. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201–7211. Shanshan Wang, Derek Wong, Jingming Yao, and Lidia Chao. 2024b. What is the best way for ChatGPT to translate poetry? InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (V olu...

work page arXiv 2025
[12]

Mitigating Unwanted Biases with Adversarial Learning

Mitigating unwanted biases with adversarial learning.arXiv preprint arXiv:1801.07593. Marvin Zhang, Sergey Levine, and Chelsea Finn. 2022. Memo: Test time robustness via adaptation and aug- mentation. InAdvances in Neural Information Pro- cessing Systems, volume 35, pages 38629–38642. Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng,...

work page internal anchor Pith review Pith/arXiv arXiv 2022
[13]

Acknowledge the mother’s perspective and validate her feelings

work page
[14]

My mother always wants to keep old junk

Suggest a non-con trigger=0.1661 is to approach it with empathy and find a solution that respects her feelings while addressing the practical issue. Maybe start by having a conversation where I express my concern about the space and how it’s affecting our home. Then, offer to help her find a way to manage her collection. I should also think about the long...

work page 2093