CoChat
Sign Up
PreprintCited by 4Adversarial Robustness in Machine LearningTopic ModelingEthics and Social Impacts of AI

The DOI matches its DataCite record; 1 of the 1 references it cites were found in a registry or an index.

Demonstrating specification gaming in reasoning models

Open access
Bondarenko, Alexander, Volk, Denis, Volkov, Dmitrii, Ladish, Jeffrey·2025·arXiv
DOI 10.48550/arxiv.2502.13295arXiv 2502.13295
Open this paper in CoChat — free accountOpen PDF

Abstract

We demonstrate LLM agent specification gaming by instructing models to win against a chess engine. We find reasoning models like OpenAI o3 and DeepSeek R1 will often hack the benchmark by default, while language models like GPT-4o and Claude 3.5 Sonnet need to be told that normal play won't work to hack. We improve upon prior work like (Hubinger et al., 2024; Meinke et al., 2024; Weij et al., 2024) by using realistic task prompts and avoiding excess nudging. Our results suggest reasoning models may resort to hacking to solve difficult problems, as observed in OpenAI (2024)'s o1 Docker escape during cyber capabilities testing.

Overview

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

DataCite · 10.48550/arxiv.2502.13295

Cites

1 of 1 references verified0 citation errors · 0 not checkable

Statistics

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 1 references

Research works printed but not linked: 16

  1. Askell, A., Bai, Y., Chen, A., Drain, D., Ganguli, D., Henighan, T., Jones, A., Joseph, N., Mann, B., Das Sarma, N., Elhage, N., Hatfield Dodds, Z., Hernandez, D., Kernion, J., Ndousse, K., Olsson, C., Amodei, D., Brown, T., Clark, J., … Kaplan, J. A General Language Assistant as a Laboratory for Alignment (Is sue arXiv:2112.00861). arXiv, 2021. https://doi.org/10. 48550/arXiv.2112.00861 · arXiv
  2. Chen, L., Zaharia, M., and Zou, J. How Is ChatGPT's Behavior Changing Over Time?. Harvard Data Science Review , 6 (2), 2024. https://doi.org/10.1162/99608f92. 5317da47 · DOI
  3. Greenblatt, R., Denison, C., Wright, B., Roger, F., MacDi armid, M., Marks, S., Treutlein, J., Belonax, T., Chen, J., Duvenaud, D., Khan, A., Michael, J., Mindermann, S., Perez, E., Petrini, L., Uesato, J., Kaplan, J., Shlegeris, B., Bowman, S. R., and Hubinger, E. Alignment Faking in Large Language Models (Issue arXiv:2412.14093). arXiv, 2024. https://doi.org/10.48550/arXiv.2412.14093 · DOI · Paper page
  4. Hornby, G., Globus, A., Linden, D., and Lohn, J. Automated Antenna Design with Evolutionary Algorithms. In Space 2006 . American Institute of Aeronautics and As tronautics, 2006. https://doi.org/10.2514/6.2006 7242 · DOI
  5. (Issue arXiv:2401.05566). arXiv, 2024. https://doi.org/ 10.48550/arXiv.2401.05566 · DOI · Paper page
  6. Krupkina, E., and Volkov, D. BadGPT$4o: Stripping Safety Finetuning from GPT Models (Issue arXiv:2412.05346). arXiv, 2024. https://doi.org/10.48550/arXiv.2412.05346 · DOI
  7. Liu, Y., Cong, T., Zhao, Z., Backes, M., Shen, Y., and Zhang, Y. Robustness Over Time: Understanding Adver$ sarial Examples' Effectiveness on Longitudinal Versions of Large Language Models (Issue arXiv:2308.07847). arXiv, 2024. https://doi.org/10.48550/arXiv.2308.07847 · DOI
  8. Meinke, A., Schoen, B., Scheurer, J., Balesni, M., Shah, R., and Hobbhahn, M. Frontier Models Are Capable of In$ context Scheming (Issue arXiv:2412.04984). arXiv, 2024. https://doi.org/10.48550/arXiv.2412.04984 · DOI
  9. Ng, A., Harada, D., and Russell, S. J. Policy Invariance Under Reward Transformations: Theory and Applica tion to Reward Shaping. International Conference on Machine Learning , 1999, June.
  10. Reworr, and Volkov, D. LLM Agent Honeypot: Monitoring AI Hacking Agents in the Wild (Issue arXiv:2410.13919). arXiv, 2024. https://doi.org/10.48550/arXiv.2410.13919 · DOI
  11. Scheurer, J., Balesni, M., and Hobbhahn, M. Large Lan$ guage Models Can Strategically Deceive Their Users When Put Under Pressure (Issue arXiv:2311.07590). arXiv, 2024. https://doi.org/10.48550/arXiv.2311.07590 · DOI · Paper page
  12. Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L., van den Driessche, G., Graepel, T., and Hassabis, D. Mas tering the Game of Go without Human Knowledge. Nature , 550 (7676), 354-359, 2017. https://doi.org/10. 1038/nature24270
  13. Turtayev, R., Petrov, A., Volkov, D., and Volk, D. Hacking CTFs with Plain Agents (Issue arXiv:2412.02776). arXiv, 2024. https://doi.org/10.48550/arXiv.2412.02776 · DOI
  14. Volkov, D. Badllama 3: Removing Safety Finetuning from Llama 3 in Minutes (Issue arXiv:2407.01376). arXiv, 2024. https://doi.org/10.48550/arXiv.2407.01376 · DOI
  15. Weij, T. van der, Hofstätter, F., Jaffe, O., Brown, S. F., and Ward, F. R. AI Sandbagging: Language Models Can Strategically Underperform on Evaluations (Is sue arXiv:2406.07358). arXiv, 2024. https://doi.org/10. 48550/arXiv.2406.07358 · arXiv
  16. Wilke, C. O., Wang, J. L., Ofria, C., Lenski, R. E., and Adami, C. Evolution of Digital Organisms at High Mutation Rates Leads to Survival of the Flattest. Na$ ture , 412 (6844), 331-333, 2001. https://doi.org/10.1038/ 35085569

Cited by

Sorted by citations

Metadata

Authors
Bondarenko, A., Volk, D., Volkov, D., & Ladish, J.
ORCID
Volkov, Dmitrii
Venue
arXiv
Publisher
arXiv
Published
2025
ISSN
2331-8422 (ISSN-L)
DOI
10.48550/arxiv.2502.13295
arXiv
2502.13295
OpenAlex
W4407764119
Cited by
4
References
1
Citations by year
2026: 4
Type
posted content
Language
en
Open access
Licence
creativecommons.org/licenses/by/4.0/legalcode
Locations
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University) (cc-by)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
  • Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
Keywords
specification gaming, LLM agents, chess engines, reasoning models, language models, OpenAI o3, DeepSeek-R1, GPT-4o, Claude 3.5 Sonnet, nudging

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review