CBX-Bench: A Human-Aligned MLLM Council for Benchmarking Concept Bottleneck Model Explanations

Concept Bottleneck Models (CBMs) are designed to make visual classification interpretable by expressing predictions through human-understandable concepts. Although interpretability is the central motivation for CBMs, they are still largely evaluated as predictive models by downstream classification accuracy, supplemented by isolated qualitative examples. This highlights a pressing need for quantitative measures, a challenge complicated by the infeasibility of ground-truth concept annotation at scale and the open nature of concept lists due to a lack of consensus. To fill this gap, we develop a multimodal large language model (MLLM) council that, given an image and its CBM explanation, produces an explanation quality score. To ground and validate the council, we first conduct a human study to establish a ground-truth reference for CBM explanation quality: for an image, annotators compare explanations from two of LF-CBM, VLG-CBM, and CBM-Suite and choose the more useful one, or mark them as equally good or equally bad, yielding 2700 judgments over 900 image-comparison items on CUB-200, ImageNet-100, and Places365. Against this human reference, our five-model council, consisting of open-weight MLLMs, recovers over 70% of strict human preference rankings, rising to 83% on items where human annotators unanimously agree. Building on this validated council, we introduce CBX-Bench, a public benchmark and leaderboard: authors of new CBMs can submit their model's explanations, and CBX-Bench scores them with the council and maintains dataset-level rankings of explanation quality. CBX-Bench thus provides a human-aligned, scalable evaluation of CBM explanations beyond accuracy and isolated qualitative examples. The benchmark is available at https://github.com/meric-karadag/cbx-bench.

Measuring Judgment Quality in Natural-Language Explanations: Evidence from Forecasting Tournaments

Decision-makers routinely rely on expert judgments accompanied by written explanations, yet explanation quality is difficult to measure at scale. Forecasting tournaments offer a natural testing ground: probabilistic judgments are paired with natural-language rationales and scored against realized outcomes. We introduce Explanation Quality Markers (EQMs), a set of sixty theory-guided reasoning patterns scored by large language models (LLMs). In a pre-registered analysis of over 55,000 forecast-rationale pairs from a multiyear forecasting tournament, EQMs predict accuracy at both the forecast and forecaster levels, consistently outperforming pre-LLM text-analysis methods. More than 90% of statistically significant pattern-level EQM-accuracy correlations match our directional hypotheses. The signal is asymmetric: EQMs identify likely underperformers more reliably than they distinguish the very best forecasters. Benchmarked against traditional indicators of forecasting skill, EQMs are the strongest predictor at the forecast level and competitive at the forecaster level, though weaker than prior accuracy. Human ratings of rationale quality are less consistently correlated with accuracy and place disproportionate weight on rationale length. Results transfer to an independent forecasting study. EQMs provide a scalable, interpretable method for extracting judgment-relevant information from written explanations.

Explanation Quality Assessment as Ranking with Listwise Rewards

We reformulate explanation quality assessment as a ranking problem rather than a generation problem. Instead of optimizing models to produce a single "best" explanation token-by-token, we train reward models to discriminate among multiple candidate explanations and learn their relative quality. Concretely, we construct per-instance candidate sets with graded quality levels and train listwise and pairwise ranking models (ListNet, LambdaRank, RankNet) to preserve ordinal structure and avoid score compression typical of pointwise regression or binary preference objectives. We observe three findings: First, ranking losses consistently outperform regression on score separation across all domains tested. Second, the optimal ranking loss depends on data characteristics: listwise objectives excel with well-separated quality tiers, while pairwise methods are more robust to noisy natural annotations. Third, when trained on carefully curated and well-structured data, small encoder models can match models that are orders of magnitude larger, suggesting that data quality matters more than model scale. Finally, when used as rewards in policy optimization, ranking-based scores enable stable convergence in settings where regression-based rewards fail entirely. Code and data are available at: https://github.com/Tankiit/PPO_Learning_to_rank