- Rating
- 1
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper introduces "adversarial psychometrics" as a pairwise protocol (a game) of problem proponents and solvers. The score is the average of the proponent score (the variance of the predictions of the other players) and the correctness scores (1 - Brier Score of the own probability predictions when a solver). This comes with experiments with two variants of the same domain (code with verifier or open-ended) with 10 modern LLMs as players of the game.
Critical strengths
Evaluating beyond human scale is becoming an important and urgent problem, but at the same time difficult.
Critical weaknesses
- "Beyond human scale". The paper talks about an absolute scale (abstract), but this is about relative (populational) evaluation scales. The paper uses the term "human scale" (title), but there's nothing about humans or anything that suggests that this goes beyond the human scale. It just ranks models, as many other games would do. A chess tournament could possibly give a similar ranking.
- Focus on Spearman, the g-factor and intelligence, when the connection with the protocol is very weak. The introduction is too generic, and starts with Spearman. This has been covered elsewhere and the connection with the paper is purely motivational. There are several (old) papers and books connecting all this, also about the g-factor and intelligence, and the use of psychometric approaches in AI evaluation. Then, including the results from the original Spearman's paper is pointless. This can be found elsewhere and more space should have been devoted to explaining the key concepts and results of the paper. Section 2 can be removed completely or moved to an appendix.
- Use and comparison with psychometric techniques is limited and biased. "Most AI benchmarks inherit this psychometric structure". This is false, most don't even extract latent factors with factor analysis, IRT, etc., and especially most of the references used in that paragraph do not use any psychometric techniques. This is simply misleading.
- Construct validity: I don't find a clear rationale why this should measure "intelligence". The hypothesis that a good intelligence discriminator has to display an important "component of intelligence" is a bold statement. Also, this doesn't clarify either whether this is a side effect (sufficient condition) of intelligence (if something is intelligent it will test intelligence well) or a necessary condition (for something to be intelligent it has to be test) or both).
- Questionable assumptions: One major methodological issue is that the paper seems to criticise that "the common issue is that the challenger possesses information unavailable to the solver". No, this is actually the natural approach for an examiner to be able to test beyond their capacity, by constructing inverse problems, such as the example of the factorisation into two primes.
- Uncertainty as variance and psychometric discrimination: "shared uncertainty among players is indicative of an uninformative challenge and should yields [sic] no separation reward" - this is actually what IRT and many other psychometric techniques do, trying to determine difficulty and discrimination from a population. Several parametric models map the information of an item with the uncertainty of the population. The connection with all this is missing, with the only psychometric technique being 100 years old (rudimentary factor analysis). The motivation that "the proposer succeeds only by finding questions that genuinely separate the audience" and targeting for variance are the traditional elements of psychometrics and the quality of items. In adaptive testing, a good evaluator is actually the one that chooses the item that better discriminates an individual or a population.
- Contribution: Making this iterative or adversarial is not new, and as the authors recognise, already explored in MathDuels.
- Lack of theoretical or empirical support for the protocol: The paper "conjectures" that variance on the probabilities discourages uninformative challenges, but this is not supported empirically or theoretically anywhere in the paper. The weakest part of the paper is that there's no theoretical justification of any of the protocols and no connection with other adversarial situations in game theory and multi-agent theory. Not even ablation of why J has to be defined as exactly the average of P and (1-B), where these two quantities (even if normalised to 1) have different sources and evolution.
- Homogeneous group of agents: The experiments are limited to a quite homogeneous group of agents (I wonder what the results would be with more diversity or even adversarial players).
- Rudimentary psychometrics: The comparison with some benchmarks and their g factor is a bit rudimentary, especially having more powerful indices of LLM capability, based on more modern psychometric techniques, such as Epoch's ECI based on IRT.
*** Some other comments or typos that may not be critical weaknesses but I think may be useful for the authors and I couldn't find a place where to put them.
Section 4 "answers A and ???" not A?
Also false commitments are left as future work, and preferences are not included.
Questions for authors
- The variance on the probabilities depend on the solvers to be calibrated. If they are not, random questions one could get more variance that meaningful questions. Can you try to use some uncalibrated models and see how this affects the results?
- Preferences could lead to variance: For instance, just asking about "red" being your preferred colour could lead to variance, but that doesn't mean that the test is measuring any intelligence (at least on the solver side), just information about the population. This means that even if for one experiment the metric converges that doesn't mean that it has to converge to any meaningful construct restricted to intelligence. All this may be artifices of the LLM population being not sufficiently diverse to be exploitable in other ways (or the prompts being limited to make things converge). Can you try with preferences or prompt the models to have some preferences and personalities and exploit this to have different answers?
- Do all the results of Figure 6 come from Artificial Analysis? Why the choice of these benchmarks and what are they measuring? For instance, AA-Omniscience is not even measuring a cognitive capability but rather the propensity of not answering a question when not known.
- Can I get more details about the models, the calibration of the models, their range in probabilities, etc. Also, to maximise the score, one model can try to regress to the population in case of doubt, and become less than perfectly calibrated could be optimal in the game. All these things should be analysed before getting the numbers from an experiments without no backing in any theory of intelligence or computational complexity.
- Given the long prompts and the complex game, good models could just get better because they understand the dynamics of the game better, rather than just because any intrinsic "intelligence" of the model. Can we clarify this?
- I may even wonder whether Astra or some other powerful models, given the equations in the prompt, have found optimal solutions to this "game" that could be incorporated as a theoretical support for this paper. I'm saying this also because GPT-6 Astra starts very high in P, in comparison with the others, even in the first round. Also because GPT-6 Astra may be using mind modelling (theory-of-mind) of the other players when the pseudonym is disclosed. Can we ask Astra to predict what the other agents are going to predict? If it performs well then the metric is loaded strongly by mind modelling (theory of mind of the other players).