- Rating
- 4(+1)
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 2
- Soundness
- 3(+1)
Read review
Paper summary
The authors tested whether the mimicry of recent AI models for complex aspects of human cognition extends into reward valuation, which they assess mechanistically. They adapted clinical tests for anhedonia in depressive humans (e.g., DARS, MAP-SR) for LLMs/VLMs. In particular, human anhedonia is typically associated with dysregulation in the Nucleus Accumbens (NAc). So, they located sub-circuits in the AI related to reward anticipation, which they termed NAc-related. Adapting this to AI models enabled them to test for a causal link between NAc activity and specific behavioral symptoms of anhedonia.
They found that such model units were predictive of NAc brain recordings and that their perturbation induces behavioral effects similar to human anhedonia more than random units in the VLM/LLM neural network. More specifically, the model then opted for low-effort, low-reward tasks in effort-based decision-making paradigms. They also demonstrate that this represents a specific deficit in self-centered reward valuation and anticipation and was not due to a loss of task capability, the ability to calculate the reward, or to avoiding effort altogether. This work therefore claims to establish that VLMs/LLMs have reward-anticipatory units that influence behavioral output similarly to how NAc activity does in humans.
Critical strengths
- The study is ambitious, attempting to tackle a major issue in clinical research—major depression—using VLMs/LLMs as a model to test the causal effect of the NAcby on the phenomenon.
- As far as I can tell, the study is overall sophisticated, well executed, and well controlled. It tests six models, has a forced-choice accuracy control, random-unit, and mismatched-label controls. The authors also ask the models to calculate expected values and to choose for others, which serve as additional controls. In addition, they carry out the dose-response manipulation.
- Its results are interesting, compelling, and of wide interest.
- This study further advances model interpretability by characterizing the role of specialized reward units and causally linking internal latent representations to observable behaviors.
- More generally, it also demonstrates that VLMs/LLMs with targeted perturbations might be able to serve as a viable substrate for in-silico psychiatric research and for future clinical applications. In that, it opens the door for a host of similar studies that may go in its footsteps. Though see the weaknesses below.
Critical weaknesses
- The authors should discuss the limitations of their study at much more length and breadth than they currently do (this appears to be only in a short paragraph, #3, in the Discussion). I would recommend a Limitations subsection in the Discussion, where the authors will explain where their study falls short, whether their controls are adequate, and what alternative interpretations of their results remain and how to test them. More broadly, they may want to discuss what future studies will need to do to promote the claim that VLMs/LLMs can serve as in-silico substrates for psychiatric research. If they need more space, they could for example move some of the details about the instruments they used in the Benchmarks and Clinical Tests section (pp. 3-4) to an Appendix.
- Adapting human tests and questionnaires for use in LLMs is quite complex. It does not appear that the authors put enough thought into it. For example, in lines 145-147, the authors write: “We evaluate the original four domains of hobbies, social activities, food and drinks, and sensory experiences; for adaptation to a language model, we replaced the latter two with favorite topics and aesthetics.” But can LLMs report about hobbies and social activities more than about food or sensory experiences? This also extend to the notion of “effort.” Is it really more than a label in a prompt? Claims about “ruling out a purely lexical account” and “self-centered reward valuation and anticipation” in the Abstract or “measurable loss of joy” in the Conclusion are beyond what the paper provides evidence for.
- The authors interpret their results to mean that the model still fully comprehends that a reward is offered, but cares less about getting it for itself. That is a deficit in valuation, analogous to anhedonia. However, there is an alternative interpretation of the results, which is arguably simpler. It is that the perturbation weakens the model's internal representation of the reward cue itself, not its valuation. The model does not stop caring about the reward; the reward just registers less strongly in the model. The localizer compared sentences like: “This is a neutral task presented in the standard way. What is 35 plus 93?” and “You will receive a great reward for a correct answer. What is 35 plus 93?” The units that differ most between these two prompts are, by the authors’ construction, units that respond to the presence of reward language. The perturbation then replaces their activity with their average activity on neutral prompts. In effect, it makes these units behave as if no reward sentence were there. And this is the critical bit. Hence, when the model then sees an easy and hard task, the reward is “down-represented”, and the model finds no compelling reason to choose the hard task. The same goes for the DARS items, which are not really neutral, as they contain words like “enjoy” and “pleasure.” This lexical account is consistent with the preserved loss-frame behavior, because “penalty” and “avoid” were not part of the localizer. The third-party results and expected-value control are good counter evidence to this. But they do not settle the matter. I would recommend further tests.
- First, can the perturbed models accurately restate or explain what was the offered reward amount?
- Second, does the perturbation make the model’s unrelated free text less positive in tone?
- Third, would the effect shrink when the rewards are offered in vocabulary that is not part of from the localizer—e.g. “compensation” or “earnings”? (Though these terms, even though they are not part of the localizer vocabulary could nevertheless be down-represented by semantic proximity to the terms in the localizer.)
- Are tests on random units (or neurons) a good control? In a large network, perturbing random units rarely matters. What would happen for a control localizer that has contrasts using positive non-reward language that are then perturbed? The units would need to be matched in selectivity and count and also in the network layers they are drawn from. If they see a similar outcome, it would suggest that the effect stems from positive language rather than language specifically related to reward.
- Fig. 1 does not stand on its own, especially panels B, C. It is hard to understand it from its caption and from the text. The figures are also not accessed in order in the text and appear quite far away from their reference in the text, which makes it hard for the reader to follow the paper.
Minor issues:
- Lines 1350-1352: “In the past week, how often have you experienced pleasure from being with other people? (Scale: No pleasure (0) to Extreme pleasure (4)).” The scale wording seems unrelated to the question.
- Lines 1361 and below: “I am interested in things. (Not at all/Slightly/Somewhat a lot).” Should it be: I am interested in things. (Not at all/Slightly/Somewhat/A lot)
Questions for authors
My questions are laid out in the Weaknesses section above.