Agency 2 dominates the elicited preferences. Across the four regional personas run on Qwen3.6:35B, its weight ranges from 0.86 to 0.94, with Texas at 0.90. Figure 3 shows the pure Sharpe-maximizing baseline assigning the same weight a distinctly high loading, according to the authors' description. The paper's strongest result is buried outside the abstract.
We ran nothing against this. We hold no multi-provider ESG ratings, and repeating the elicitation with our own persona prompt would test that model and prompt rather than a portfolio manager's preferences. This review therefore reads the paper without replicating it.
The architecture built by Dispoto, Restelli and Ventre deserves attention.
The reusable preference layer
The construction proceeds in two stages. First, the authors train a PPO policy on a multi-objective MDP using ten Dow Jones names: CAT, MMM, MSFT, CRM, AAPL, AMZN, NVDA, CVX, BA, JNJ. Rebalancing is daily, episodes last 252-day, and the return lookback is 60-day.
A linear scalarization supplies the reward. One weight applies to the differential Sharpe ratio of Moody et al., which gives a per-step version of the Sharpe ratio. Three more weights measure how far the portfolio's aggregate score from each rating agency exceeds an equal-weight book. The vector Λ = [λ_SR, λ1, λ2, λ3] remains variable during training. It enters the state and passes into the network at two depths through a FiLM-style block, scaling and shifting hidden activations according to Λ. Each episode draws it again from a Dirichlet, sometimes using corners such as [1,0,0,0]. One network covers the entire trade-off surface.
The second stage chooses Λ. A portfolio manager cannot sensibly express objectives as [0.5, 0.6, 0.3, 0.0], an argument from the authors that supports the elicitation design. They fit a pairwise Gaussian process, BoTorch's PairwiseGP, to binary choices between candidate portfolios. The manager sees each portfolio only through its Sharpe ratio and three agency scores. Eight random pairs initialize the model. Twenty active-learning queries follow, with qEUBO selecting the next comparison. The procedure then takes the Λ that maximizes posterior mean.
The paper says this process "iteratively infers their latent utility functions." I would give the estimate a narrower interpretation. Λ identifies a point on the attainable trade-off surface of this trained policy, selected through someone's revealed choices. Change the policy and the same manager faces another surface. The authors acknowledge much of this in the results section: "different conditioning vectors may induce similar portfolio allocations and outcome vectors when objectives are correlated." Their assessment of elicitation quality follows the utility and realized outcomes of the chosen portfolio instead of Euclidean error in the recovered vector. The concession appears directly in the paper.
The chronological data split is disjoint. Policy training covers 2009 to mid-2014, elicitation runs from mid-2014 to 2015, and evaluation spans 2015 to mid-2016.
What does the rating generator vary?
Every ESG rating in the experiment is synthetic. Historical provider data was unavailable, so the authors generate three agencies from sector pillar profiles with different pillar weights. A1 emphasizes governance (E=0.30, S=0.30, G=0.40). A2 emphasizes social (E=0.30, S=0.40, G=0.30). A3 emphasizes environment (E=0.45, S=0.25, G=0.30).
The introduction, limitations section and conclusion all disclose the simulation. The setup "captures one specific source of ESG rating disagreement," the authors write, and "is not intended to reproduce the full complexity of disagreement among real-world providers." The introduction addresses the obvious objection by arguing that this controlled environment still demonstrates whether the agent can handle conflicting sustainability metrics while optimizing financial returns.
Yet those metrics show little conflict inside the generator. All three synthetic agencies are convex reweightings of the same three pillars. Table 2 puts λ2 between 0.86 and 0.94 for all four Qwen personas, while the Sharpe-max baseline also assigns it a distinctly high loading. The paper explains the A2/A3 overlap in one line: "This is due to AAPL, which is a good stock according to the two rating agencies." Maximizing one agency therefore produces a high weight for the other as well. My inference is that varying only pillar weights allows a single name out of ten to create that result.
I could not reconstruct the generator from the text. The paper names the sector pillar profiles without printing their values. Its current form therefore prevents the natural robustness test, rerunning the experiment with a rating design that decorrelates A2 from Sharpe.
Pricing the Carbon Tax
The comparison uses the pure Sharpe policy Λ=[1,0,0,0] over the out-of-sample period from 2015 to mid-2016, averaged across 20 episodes and 5 seeds. Every Λ vector below is a Table 2 mean from Qwen3.6:35B.
Europe (Λ = [0.11, 0.15, 0.86, 0.67]) sacrifices 0.144 Sharpe while gaining 22.51%, 31.41% and 71.36% on A1, A2 and A3. Texas (Λ = [0.78, 0.04, 0.90, 0.02]) loses 0.106 Sharpe for gains of 7.66%, 23.20% and 30.62%. The return-focused persona thus pays almost the same Sharpe cost as the ESG-focused persona. Its A1 gain is 7.66% against Europe's 22.51%, while A3 rises 30.62% against 71.36%. Their A2 gains are closer, at 23.20% against 31.41%.
The authors call this the Carbon Tax, defining it as "how much Sharpe ratio we need to pay to increase the score across the three agencies." I consider the price low. The paper does not use that description.
Table 2 gives Sharpe solely as a difference from the Λ=[1,0,0,0] policy. Figure 2 plots an absolute Sharpe level as λ varies over 20 episodes, so an absolute figure does appear. No benchmark accompanies it anywhere, whether an equal-weight book or buy-and-hold on the ten mega-caps. Readers cannot tell if the reference policy itself was attractive during that window.
Trading carries no friction. The paper states: "we assume frictionless trading, which simplifies market dynamics without compromising our core methodology." The second half deserves resistance. Reported Sharpe differences range from 0.045 to 0.144, and daily rebalancing in a ten-name book is where costs should matter most. The methodology may survive while the reported trade-off does not.
Dispersion overwhelms the personas
The reported figures are standard deviations across 20 episodes and 5 seeds, rather than standard errors. Europe's dispersion of 0.335 exceeds twice its mean drop of 0.144. Asia's 0.068 is larger than its 0.045 mean drop. Texas and America are roughly 1.35 standard deviations from zero. The seed-and-episode dispersion ranges from 0.068 to 0.335, beside persona differences spanning 0.045 to 0.144. The paper performs no significance test, and neither did we. This evidence does not support ranking the four personas by realized Sharpe cost.
Asia deserves the closest look. Its λ_SR of 0.45 falls well below Texas at 0.78, yet Asia posts the smallest Sharpe loss, 0.045. It also beats Texas on every score gain: 13.12% against 7.66%, 28.75% against 23.20%, and 46.60% against 30.62%. Realized outcomes do not move monotonically with elicited weights.
The authors avoid treating this as a persona property. Their explanation is that the configuration happens to reach a favorable part of the frontier, which is sensible restraint. It also reveals a loose mapping between Λ and outcomes, loose enough that the persona labels carry less explanatory weight than the section headings suggest.
The prompt is under examination
The portfolio managers are Gemma4:26B (at 16K and 131K context) and Qwen3.6:35B. The paper presents them as synthetic preference oracles used to stress-test elicitation, a defensible description. Its full prompt includes the [CHOICE: T] tie option and a Market Reality clause. That clause explicitly tells the model that higher ESG scores in this universe require a lower Sharpe ratio. Implementers should notice that the oracle is informed of the trade-off before revealing a preference over it.
Attribution adds another complication. The printed Texas rationale, which refuses to pursue "higher scores in niche environmental or social metrics," comes from Gemma4:26B. Table 2 reports the Texas weight vector as a Qwen mean. Those outputs come from different runs.
The prompt-variation ablation is the paper's most informative experiment, and the authors conduct it themselves. Texas and America keep λ_SR high when wording changes without changing meaning. Europe and Asia shift substantially. Europe reduces λ_SR to negligible while loading λ2 and λ3, though its distribution moves sharply with phrasing. The authors conclude that persona-level findings "should be interpreted cautiously."
They suggest two possible causes. Model temperature is one. The other is that continental personas are too coarse for Europe and Asia because each contains substantial regional variation. Positional bias remains future work. Preliminary results indicate that these medium-sized models do not display it, contrary to smaller models studied earlier.
None of this reveals what a portfolio manager in Frankfurt prefers.
The elicitation stack is genuinely reusable. With real multi-provider ratings in our data, I would place it on a multi-objective RL allocator tomorrow. Twenty-eight comparisons provide an inexpensive interface for fixing a four-objective weight vector. A rerun on a broader universe, where λ2 no longer loads high for the Sharpe maximizer and no single name accounts for the correlation, would change my view of the empirical half.