Speakable Welfare
Muhammad Zane Abdullah, Matthew Elliott, Anamaria Leonescu, Sharan Nagarajan, Parivrudh Rajeev Sharma · Team Latent Minds
Submitted to Digital Minds Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
A model’s self-report is an increasingly attractive way to monitor its internal states such as goals, preferences, or welfare, but naturally such reports are hard to interpret if the underlying state is not accessible through the model’s language-generating mechanisms. We address this question for an internal direction associated with better or poorer performance during reinforcement learning (RL) experiments, i.e. a functional welfare state. We use a Jacobian lens to measure speakability, i.e. the extent to which an internal direction is represented in the model’s verbalizable subspace, and compare trained welfare directions with random, pre-RL, and naive controls. our results suggest that welfare-related representations can be present and partly speakable before RL, with RL selectively changing how strongly they are represented
Reviews
The study builds on results by Han et al. (2026), who “extract concept vectors for rewarded and punished trajectories”. The current study asks, with respect to this direction in activation space: “does it enter the model’s J-space, and does this affect its self-report?” (p. 2).
The presentation of the results is highly technical; it is hard to follow for readers like me (who are not directly involved this kind of research). Given that the paper also has a detailed and long appendix with two parts, I wonder if the main text could have focused on a less technical presentation of the methods and results. In particular, key terms like “speakable” are never defined (or the definition in the introduction is incomplete: “However, for a state to be speakable; represented in the subspace with privileged influence on output tokens or J-space (Gurnee et al. 2026.).”).
If I understand correctly, the main answer to the research question is “yes, but the welfare axis already exists in the J-space prior to RL”. I could not find an explicit answer to the second part of the question, but it seems there is no direct influence on self-report. I would recommend making the answers to the research question more explicit.
Read full reviewShow less
This project takes on an important and natural question: to what extent are internal representations associated with welfare present in the J-space, and thus verbalisable? This is an excellent topic, but I found that the results were somewhat obscured in the paper, which is long, seems to be largely LLM-written, and not easy to follow.
Due to severe time constraints, this review may contain mistakes or oversights. For the same reason, it focuses on the paper’s key idea, not the detailed execution: The paper combines two techniques from prior research in an interesting way; at the same time, I am not sure what the theoretical significance of this is. One worry is that their interpretation of the J-Space "presence in the J-space is necessary for a report to be grounded in the state it describes“ is precisely what’s undermines by David Chalmers’ critique that they cite earlier. I also found the discussion section hard to follow.
Congrats on thihs really strong project, and already working with the freshly-released J-Lens methos. Love the falsification logic (presence in the J-space doesn't validate a self-report but absence falsifies it).
Great methodology, too!
Two nitpicks:
1) the cited pre-registration commit isn't in the public history
2) And I'd recommend another editing pass to clean up slop wording (e.g. the headline 'the gap' set of my slop alarm) and some broken sentences.
Cite this project
@misc{abdullah2026speakable,
title = {{Speakable Welfare}},
author = {Muhammad Zane Abdullah and Matthew Elliott and Anamaria Leonescu and Sharan Nagarajan and Parivrudh Rajeev Sharma},
year = {2026},
month = aug,
note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/speakable-welfare-mt4y}},
url = {https://apartresearch.com/sprints/projects/speakable-welfare-mt4y}
}More from Digital Minds Research Sprint
- 1st placeView project: Readable but Not Causal: Limits of Self-Attributed Welfare Representations in Language Models
Readable but Not Causal: Limits of Self-Attributed Welfare Representations in Language Models
Welfare-like internal representations are increasingly studied as candidate evidence about AI systems. Their entity attribution—whether a valence state belongs to the active assistant or to a merely represented other—is …
- 2nd placeView project: Project Anchored
Project Anchored
Team Wagner
Anchoring vignettes are the standard survey-methodology fix for self-reports that are not comparable across respondents. This project applies them to language models for the first time, using code generation as a …
- 3rd placeView project: Model, Instance, or Persona? Measuring Affective Signals in Public Text After an AI Is Retired
Model, Instance, or Persona? Measuring Affective Signals in Public Text After an AI Is Retired
This sprint asks whether the assistant identifies as a model, an instance, or a persona. I ask which of the three its users name. When a company retires an AI model, users write about the loss in public, and what they …