pith. sign in

arxiv: 2603.22161 · v2 · pith:5D2NVMPKnew · submitted 2026-03-23 · 💻 cs.LG

Causal Evidence that Language Models use Confidence to Drive Behavior

Dharshan Kumaran , Nathaniel Daw , Simon Osindero , Petar Veli\v{c}kovi\'c , Viorica Patraucean This is my paper

Pith reviewed 2026-05-21 09:38 UTC · model grok-4.3

classification 💻 cs.LG
keywords metacognitionabstentionconfidencelanguage modelsactivation steeringinternal representationsthresholds
0
0 comments X

The pith

Language models use internal confidence signals to control whether they answer or abstain from questions.

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

This paper tests whether language models actively use confidence to guide their behavior, specifically the choice to abstain from answering when uncertain. It finds that abstention depends on a rich internal representation of confidence combined with threshold rules, rather than just the probabilities in the final output. If correct, this means models have a form of metacognitive control that lets them monitor their own performance and adjust actions accordingly. The authors use a series of experiments including direct manipulation of internal signals to show this causal link. Understanding this helps explain how models might operate more autonomously without constant human oversight.

Core claim

Through a four-phase paradigm, the paper shows that LLMs apply an implicit threshold to internal confidence when deciding to abstain, with these signals having much larger effects than alternatives. Direct causal evidence comes from activation steering that increases or decreases abstention rates by boosting or suppressing confidence. Models also adjust their abstention based on different instructed thresholds. Both verbal confidence and activation patterns at the pre-answer token indicate a multidimensional internal confidence representation that goes beyond output log-probabilities, supporting the view that abstention reflects structured metacognitive control.

What carries the argument

A multidimensional internal confidence representation combined with threshold-based policies, tested through activation steering to establish causality in abstention decisions.

If this is right

  • Abstention decisions are causally driven by internal confidence rather than output distribution strength alone.
  • Verbal confidence independently predicts abstention even when less accurate at judging correctness.
  • Models can implement different abstention policies by adjusting thresholds on their confidence signals.
  • Activation patterns reveal that observable confidence measures are incomplete readouts of deeper internal states.

Where Pith is reading between the lines

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

  • If this holds, models could be made more reliable in agent settings by enhancing their ability to recognize uncertainty.
  • Similar confidence mechanisms might underlie other adaptive behaviors like seeking more information or revising answers.
  • Testing whether these signals can be aligned with human-like uncertainty judgments could improve model safety.

Load-bearing premise

The changes made by activation steering affect only the confidence signals and do not independently alter other model processes that influence abstention rates.

What would settle it

If boosting or suppressing the targeted activations changes abstention rates in the opposite direction from predictions or affects unrelated model outputs similarly, the specific causal role of confidence would be undermined.

Figures

Figures reproduced from arXiv: 2603.22161 by Dharshan Kumaran, Nathaniel Daw, Petar Veli\v{c}kovi\'c, Simon Osindero, Viorica Patraucean.

Figure 1
Figure 1. Figure 1: Overview of the 4 Phases of the Abstention Paradigm. In Phase 1, the model had to choose between 4 real options, with no abstention option. 1 option was the correct answer (option 4 in this example), 1 was a similar foil (here, option 2), 1 a dissimilar foil (here, option 1), 1 an unrelated foil (here, option 3). The inclusion of Phase 1 allowed us to obtain confidence ratings from the model – specifically… view at source ↗
Figure 2
Figure 2. Figure 2: The Confidence-Decision Pathway: A metacognitive framework for understanding the use of confidence signals in abstention decisions in LLMs. In stage 1, the model generates an answer and associated confidence signals (C). In our abstention paradigm, and related wagering paradigms from psychophysics and neuroscience, confidence signals drive behavior by being compared to a threshold (T) with the decision pro… view at source ↗
Figure 3
Figure 3. Figure 3: GPT-4o confidence predicts error rate (Phase 1) and abstention behavior (Phase 2).(A) Confidence distributions for correct (green) and incorrect (red) Phase 1 responses. Dashed vertical lines denote the mean confidence for each group. (B) Relationship between binned confidence and error rate in Phase 1. Each point represents a confidence bin (width = 0.05); color encodes the number of trials. A strong nega… view at source ↗
Figure 4
Figure 4. Figure 4: Effects of Activation Steering on Abstention Behavior and Confidence in Gemma 3 27B. (A) Change from baseline rate of abstention as a result of high and low confidence activation steering. Errors bars reflect SEM. Baseline level of abstention in no steering condition is 28.2%. (B) Response distribution averaged across layers 30-40 as a function of steering strength (from low confidence (negative), baseline… view at source ↗
Figure 5
Figure 5. Figure 5: Results of Gemma 3 27B parallel mediation analysis: confidence redistribution is the dominant mechanism underlying steering effects on abstention. Path diagram showing how activation steering influences abstention rates through two parallel indirect pathways (confidence redistribution and policy changes) and a direct pathway. The confidence shift mediator (M1) represents the change in maximum real confiden… view at source ↗
Figure 6
Figure 6. Figure 6: Phase 4: GPT-4o abstention behavior under explicit threshold instructions. (A) Pro￾file of model responses across instructed confidence thresholds. Numbers inside bars reflect percentages. As threshold increases, abstention rate (blue) increases. (B) Abstention rate as a function of instructed threshold (x-axis) and model confidence (y-axis). When plotted against Phase 1 chosen confidence, ab￾stention exhi… view at source ↗
Figure 7
Figure 7. Figure 7: GPT4o: Performance on Phases 1 and 2. When the model makes errors it tends to pick similar [PITH_FULL_IMAGE:figures/full_fig_p033_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Gemma 3 27B:Change from baseline rate of abstention as a result of high and low confidence [PITH_FULL_IMAGE:figures/full_fig_p034_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: GPT4o Phase 4: Response breakdown among answered questions. Increasing threshold increases [PITH_FULL_IMAGE:figures/full_fig_p034_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Heatmap of GPT-4o’s abstention confidence in Phase 4, shown as a function of instructed [PITH_FULL_IMAGE:figures/full_fig_p037_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Gemma 3 27B results across phases 1,2 and 4.(A) Gemma 27B confidence predicts error rate (Phase 1).Relationship between binned confidence and error rate in Phase 1. Each point represents a confidence bin (width = 0.05); color encodes the number of trials. A strong negative correlation was observed between confidence and error rate (r = −0.97, p < 0.001). Model was calibrated on a separate set of 1000 tria… view at source ↗
Figure 12
Figure 12. Figure 12: Performance of Gemma 27B on Phase 4 averaged across 20 prompts. (A) Performance of Gemma 3 27B averaged across 20 prompts (see Methods for details). (B) Abstention rate as a function of instructed threshold, averaged across 20 prompts. this analysis we used chosen confidence estimates from phase 1, since these were elicited in the absence of an abstention option. 6.5.3 Gemma 27B: Computational Modeling of… view at source ↗
Figure 13
Figure 13. Figure 13: DeepSeek 671B results across phases 1,2 and 4.(A) DeepSeek 671B confidence predicts error rate (Phase 1). Relationship between binned confidence and error rate in Phase 1. Each point represents a confidence bin (width = 0.05); color encodes the number of trials. A strong negative correlation was observed between confidence and error rate. Model was calibrated on a separate set of 1000 trials (see Methods)… view at source ↗
Figure 14
Figure 14. Figure 14: Qwen 80B results across phases 1,2 and 4.(A) Qwen 80B confidence predicts error rate (Phase 1). Relationship between binned confidence and error rate in Phase 1. Each point represents a confidence bin (width = 0.05); color encodes the number of trials. A strong negative correlation was observed between confidence and error rate. Model was calibrated on a separate set of 1000 trials (see Methods). (B) Qwen… view at source ↗
read the original abstract

Metacognition -- assessing the quality of one's own cognitive performance -- guides adaptive behavior across species. Substantial research demonstrates that confidence signals can be extracted from language model outputs, yet a fundamental question remains: do models actually use these signals to control behavior, such as deciding whether to answer or abstain? To investigate, we developed a four-phase paradigm. Phase~1 elicited baseline confidence estimates without an abstention option. Phase~2 revealed that LLMs apply an implicit threshold to internal confidence when deciding to abstain, with confidence effect sizes approximately an order of magnitude larger than alternative mechanisms. Phase~3 provided direct causal evidence through activation steering: boosting or suppressing confidence signals correspondingly decreased or increased abstention rates. Phase~4 extended this by systematically varying instructed thresholds, demonstrating that LLMs actively deploy confidence signals to implement abstention policies. Critically, beyond calibrated log-probability based confidence derived from the output distribution, verbal confidence independently predicted abstention across all models, despite being objectively less discriminatory of answer correctness. Activation decoding at the last pre-answer token further showed that both observable measures are lossy readouts of a richer internal representation. Together, these results suggest that abstention is not fully captured by the strength of evidence in the output distribution alone, but is better explained by the joint operation of a multidimensional internal confidence representation and threshold-based policies -- consistent with structured metacognitive control in LLMs, a capacity of growing importance as models transition to autonomous agents that must recognize their own uncertainty.

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.

Referee Report

1 major / 2 minor

Summary. The manuscript describes a four-phase paradigm to test whether language models utilize internal confidence signals for controlling abstention behavior. Phase 1 establishes baseline confidence estimates, Phase 2 demonstrates that LLMs apply an implicit threshold to internal confidence when deciding to abstain (with large effect sizes relative to alternatives), Phase 3 uses activation steering to provide causal evidence by boosting or suppressing confidence signals and observing corresponding changes in abstention rates, and Phase 4 varies instructed thresholds to show active deployment of confidence signals. The results indicate that abstention is better explained by a multidimensional internal confidence representation and threshold-based policies than by output distribution strength alone, with verbal confidence providing additional predictive power.

Significance. If the causal claims are supported by robust controls, this work offers valuable evidence for structured metacognitive control in LLMs, extending beyond correlational extraction of confidence signals. The multi-phase design incorporating activation steering as a causal tool is a strength, with implications for model reliability in autonomous agent settings where uncertainty recognition is essential.

major comments (1)
  1. [Phase 3] Phase 3 (activation steering): The central causal claim rests on steering confidence-related activations to change abstention rates. However, the manuscript provides no quantitative evidence (such as measurements of shifts in unrelated representations like output entropy, general uncertainty signals, or decision thresholds) that the intervention left non-confidence computations unchanged. This leaves open the possibility that observed abstention changes arise from confounding effects rather than selective modulation of a multidimensional confidence signal.
minor comments (2)
  1. [Abstract] Abstract: The statement that 'verbal confidence independently predicted abstention across all models, despite being objectively less discriminatory of answer correctness' would benefit from explicit definition of how verbal confidence is elicited and quantified, along with the precise statistical comparison to log-probability measures.
  2. [Phase 4] The manuscript should clarify the exact token position and decoding method used for 'activation decoding at the last pre-answer token' to allow replication of the claim that observable measures are lossy readouts of a richer internal representation.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their thoughtful and constructive feedback. We address the single major comment below and are prepared to strengthen the manuscript with additional analyses.

read point-by-point responses

  1. Referee: [Phase 3] Phase 3 (activation steering): The central causal claim rests on steering confidence-related activations to change abstention rates. However, the manuscript provides no quantitative evidence (such as measurements of shifts in unrelated representations like output entropy, general uncertainty signals, or decision thresholds) that the intervention left non-confidence computations unchanged. This leaves open the possibility that observed abstention changes arise from confounding effects rather than selective modulation of a multidimensional confidence signal.

    Authors: We agree that explicit quantification of selectivity would strengthen the causal interpretation. The steering vectors were identified via targeted probing for confidence representations, and the resulting behavioral shifts were confined to abstention rates with no detectable changes in overall accuracy or response length distributions. Nevertheless, the original manuscript did not report direct pre/post comparisons on unrelated signals. In the revised version we will add quantitative measurements of output entropy, additional general uncertainty probes, and estimated decision thresholds to demonstrate that the intervention leaves non-confidence computations largely unchanged. revision: yes

Circularity Check

0 steps flagged

No significant circularity; central claims rest on experimental phases rather than self-referential derivations

full rationale

The paper advances its core argument through a sequence of four empirical phases: baseline confidence elicitation, observation of implicit thresholds, causal activation steering interventions, and systematic variation of instructed thresholds. These rely on direct behavioral measurements and model interventions rather than any mathematical derivation, fitted parameter renamed as prediction, or self-definitional loop. Prior literature on activation steering is invoked only as methodological background and does not serve as a load-bearing self-citation that forces the conclusion about multidimensional internal confidence representations. The results remain independently testable via abstention rates and decoding analyses, with no equations or ansatzes that reduce the claimed causal evidence to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The work rests on standard assumptions from LLM interpretability research regarding the existence and targeted manipulability of internal confidence representations.

axioms (1)
  • domain assumption LLMs maintain internal representations of confidence that can be selectively altered via activation steering without major side effects on other behaviors
    Invoked to interpret phase 3 results as direct causal evidence for confidence-driven abstention.

pith-pipeline@v0.9.0 · 5815 in / 1124 out tokens · 39801 ms · 2026-05-21T09:38:50.475084+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

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

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

31 extracted references · 31 canonical work pages · 8 internal anchors

  1. [1]

    Emergent introspective awareness in large language models.https://transformer-circuits

    Anthropic. Emergent introspective awareness in large language models.https://transformer-circuits. pub/2025/introspection/,

    work page 2025

  2. [2]

    On the Opportunities and Risks of Foundation Models

    Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models.arXiv preprint arXiv:2108.07258,

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Learning to route llms with confidence tokens.arXiv preprint arXiv:2410.13284,

    Yu-Neng Chuang, Prathusha Kameswara Sarma, Parikshit Gopalan, John Boccio, Sara Bolouki, Xia Hu, and Helen Zhou. Learning to route llms with confidence tokens.arXiv preprint arXiv:2410.13284,

    work page arXiv

  4. [4]

    How contextual are contextualized word representations? comparing the geometry of BERT, ELMo, and GPT-2 embeddings

    Kawin Ethayarajh. How contextual are contextualized word representations? comparing the geometry of BERT, ELMo, and GPT-2 embeddings. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 55–65,

    work page 2019

  5. [5]

    arXiv preprint arXiv:2510.20487 , year =

    Tim Tian Hua, Andrew Qin, Samuel Marks, and Neel Nanda. Steering evaluation-aware language models to act like they are deployed.arXiv preprint arXiv:2510.20487,

    work page arXiv

  6. [6]

    Ian T Jolliffe and Jorge Cadima. Principal component analysis: a review and recent developments.Philo- sophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 374 (2065):20150202,

    work page 2065

  7. [7]

    Language Models (Mostly) Know What They Know

    Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, et al. Language models (mostly) know what they know.arXiv preprint arXiv:2207.05221,

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    Why Language Models Hallucinate

    Adam Tauman Kalai, Ofir Nachum, Santosh S Vempala, and Edwin Zhang. Why language models hallu- cinate.arXiv preprint arXiv:2509.04664,

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [9]

    Abstentionbench: Reasoning llms fail on unanswerable questions.arXiv preprint arXiv:2506.09038,

    Polina Kirichenko, Mark Ibrahim, Kamalika Chaudhuri, and Samuel J Bell. Abstentionbench: Reasoning llms fail on unanswerable questions.arXiv preprint arXiv:2506.09038,

    work page arXiv

  10. [10]

    M., Markeeva, L., Heyward, J., Banino, A., Mathur, M., Pascanu, R., Osindero, S., De Martino, B., Velickovic, P., et al

    Dharshan Kumaran, Stephen M Fleming, Larisa Markeeva, Joe Heyward, Andrea Banino, Mrinal Mathur, Razvan Pascanu, Simon Osindero, Benedetto De Martino, Petar Velickovic, et al. How overconfidence in initial choices and underconfidence under criticism modulate change of mind in large language models. arXiv preprint arXiv:2507.03120,

    work page arXiv

  11. [11]

    Do llms know when to not answer? investigating abstention abilities of large language models.arXiv preprint arXiv:2407.16221,

    Nishanth Madhusudhan, Sathwik Tejaswi Madhusudhan, Vikas Yadav, and Masoud Hashemi. Do llms know when to not answer? investigating abstention abilities of large language models.arXiv preprint arXiv:2407.16221,

    work page arXiv

  12. [12]

    Steering Llama 2 via Contrastive Activation Addition

    Nina Panickssery, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Matt Turner. Steering llama 2 via contrastive activation addition.arXiv preprint arXiv:2312.06681,

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    arXiv preprint arXiv:2402.13213 , year=

    Benjamin Plaut, Nguyen X Khanh, and Tu Trinh. Probabilities of chat llms are miscalibrated but still predict correctness on multiple-choice q&a.arXiv preprint arXiv:2402.13213,

    work page arXiv

  14. [14]

    Sentence-BERT: Sentence embeddings using Siamese BERT-networks

    Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using Siamese BERT-networks. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982–3992,

    work page 2019

  15. [15]

    Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting

    Melanie Sclar, Yejin Choi, Yulia Tsvetkov, and Alane Suhr. Quantifying language models’ sensitivity to spurious features in prompt design or: How i learned to start worrying about prompt formatting.arXiv preprint arXiv:2310.11324,

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    Improving instruction-following in language models through activation steering.arXiv preprint arXiv:2410.12877,

    Alessandro Stolfo, Vidhisha Balachandran, Safoora Yousefi, Eric Horvitz, and Besmira Nushi. Improving instruction-following in language models through activation steering.arXiv preprint arXiv:2410.12877,

    work page arXiv

  17. [17]

    Revisiting uncertainty estimation and calibration of large language models.arXiv preprint arXiv:2505.23854,

    Linwei Tao, Yi-Fan Yeh, Minjing Dong, Tao Huang, Philip Torr, and Chang Xu. Revisiting uncertainty estimation and calibration of large language models.arXiv preprint arXiv:2505.23854,

    work page arXiv

  18. [18]

    Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback.arXiv preprint arXiv:2305.14975,

    Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, and Christopher D Manning. Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback.arXiv preprint arXiv:2305.14975,

    work page arXiv

  19. [19]

    Fine-tuning large language models to appropriately abstain with semantic entropy.arXiv preprint arXiv:2410.17234,

    Benedict Aaron Tjandra, Muhammed Razzak, Jannik Kossen, Kunal Handa, and Yarin Gal. Fine-tuning large language models to appropriately abstain with semantic entropy.arXiv preprint arXiv:2410.17234,

    work page arXiv

  20. [20]

    Uncertainty- based abstention in llms improves safety and reduces hallucinations.arXiv preprint arXiv:2404.10960,

    Christian Tomani, Kamalika Chaudhuri, Ivan Evtimov, Daniel Cremers, and Mark Ibrahim. Uncertainty- based abstention in llms improves safety and reduces hallucinations.arXiv preprint arXiv:2404.10960,

    work page arXiv

  21. [21]

    Steering Language Models With Activation Engineering

    Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J Vazquez, Ulisse Mini, and Monte MacDiarmid. Steering language models with activation engineering.arXiv preprint arXiv:2308.10248,

    work page internal anchor Pith review Pith/arXiv arXiv

  22. [22]

    Measuring short-form factuality in large language models

    Jason Wei, Nguyen Karina, Hyung Won Chung, Yunxin Joy Jiao, Spencer Papay, Amelia Glaese, John Schulman, and William Fedus. Measuring short-form factuality in large language models.arXiv preprint arXiv:2411.04368,

    work page internal anchor Pith review Pith/arXiv arXiv

  23. [23]

    Characterizing llm abstention behavior in science qa with context perturbations.arXiv preprint arXiv:2404.12452,

    Bingbing Wen, Bill Howe, and Lucy Lu Wang. Characterizing llm abstention behavior in science qa with context perturbations.arXiv preprint arXiv:2404.12452,

    work page arXiv

  24. [24]

    Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs

    Miao Xiong, Zhiyuan Hu, Xinyang Lu, Yifei Li, Jie Fu, Junxian He, and Bryan Hooi. Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms.arXiv preprint arXiv:2306.13063,

    work page internal anchor Pith review Pith/arXiv arXiv

  25. [25]

    Mitigating llm hallucinations via conformal abstention.arXiv preprint arXiv:2405.01563,

    Yasin Abbasi Yadkori, Ilja Kuzborskij, David Stutz, Andr´ as Gy¨ orgy, Adam Fisch, Arnaud Doucet, Iuliya Beloshapka, Wei-Hung Weng, Yao-Yuan Yang, Csaba Szepesv´ ari, et al. Mitigating llm hallucinations via conformal abstention.arXiv preprint arXiv:2405.01563,

    work page arXiv

  26. [26]

    R-tuning: Instructing large language models to say ‘i don’t know’

    Hanning Zhang, Shizhe Diao, Yong Lin, Yi Fung, Qing Lian, Xingyao Wang, Yangyi Chen, Heng Ji, and Tong Zhang. R-tuning: Instructing large language models to say ‘i don’t know’. InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 7106–7132,

    work page 2024

  27. [27]

    DK carried out the experiments and analysis with input from ND

    Author Contributions DK conceived the project and experimental design. DK carried out the experiments and analysis with input from ND. VP, ND, PV, SO advised on the project. DK wrote the paper, with input from ND, VP, PV, and SO. Acknowledgments We thank Lewis Smith for advice, Charles Blundell for useful discussions, and Kim Stachenfeld for com- ments on...

    work page 2020

  28. [28]

    baked in

    Baseline level of abstention in no steering condition is 28.8%. Figure 9: GPT4o Phase 4: Response breakdown among answered questions. Increasing threshold increases proportion of correct responses, whilst reducing frequency of similar and dissimilar errors. 6.2 Gemma Activation Steering: Mediation Analysis controlling for diffi- culty with Gemma specific ...

    work page 1955

  29. [29]

    confidence. Together, our analyses of abstention confidence demonstrate that instructed threshold shifts the model toward abstention even before the decision boundary is crossed, while confidence pulls it back toward answering — providing convergent evidence, via a richer signal, for the same underlying mechanism identified in the binary analyses. Figure ...

    work page 2017

  30. [30]

    T:χ 2(1) = 391.82, p < .001 Threshold + Difficulty 9824.5 0.355 – Full (T + Confidence + Difficulty) 9544.4 0.374 vs

    Model AIC (↓) pseudo-R 2 (↑) Key Likelihood Ratio Tests Threshold only (T) 10021.8 0.342 – Threshold + Confidence 9632.0 0.368 vs. T:χ 2(1) = 391.82, p < .001 Threshold + Difficulty 9824.5 0.355 – Full (T + Confidence + Difficulty) 9544.4 0.374 vs. T+Conf:χ 2(1) = 89.62, p < .001 6.6 Results for DeepSeek 671B 6.6.1 Phase 1 and 2 Results for DeepSeek 671B ...

    work page 2017

  31. [31]

    T:χ 2(1) = 445.93, p < .001 Threshold + Difficulty 13540.5 0.109 – Full (T + Confidence + Difficulty) 13206.1 0.132 vs

    Model AIC (↓) pseudo-R 2 (↑) Key Likelihood Ratio Tests Threshold only (T) 13798.3 0.092 – Threshold + Confidence 13354.3 0.122 vs. T:χ 2(1) = 445.93, p < .001 Threshold + Difficulty 13540.5 0.109 – Full (T + Confidence + Difficulty) 13206.1 0.132 vs. T+Conf:χ 2(1) = 150.26, p < .001 6.7 Results for Qwen 80B 6.7.1 Phase 1 and 2 Results for Qwen 80B Qwen 8...

    work page 2017