개념 병목 모델(CBM)의 사전 정의된 개념 집합 의존성과 정보 누수 문제를 해결하기 위해, LMM이 생성한 캡션을 병목층으로 활용해 텍스트만으로 분류를 수행하는 CaBM(Caption Bottleneck Models) 프레임워크를 제안한다. CaBM은 설계상 누수 없는 구조를 보장하면서 학습 후 텍스트 분류기 분석을 통해 데이터셋 특화 개념을 자율적으로 발견하며, 세밀 및 조대 벤치마크에서 경쟁력 있는 정확도와 해석 가능성을 동시에 달성한다.
Many high-stakes screening decisions in finance amount to predicting a rare outcome from unstructured text, such as whether a founder will successfully build a large company, or which crowdfunded loan will fail to attract funding. When positives are this rare, simple models may be preferable to more complex ones. For example, a model that learns a weight for each feature may have too little signal to estimate those weights reliably, and also inspecting the model means reading a list of fitted coefficients rather than a simple rule that can be checked by eye. We introduce Rule Forest, which uses a large language model to generate simple YES/NO questions, rather than using it to directly predict outcomes, and combines the strongest questions into an equal-weight "green-flags" scorecard. This scorecard is as accurate as an equivalent model that learns a weight per question, within noise, and its advantage appears largest when positives are rarest. The result is an interpretable list of rules that a domain expert can edit and extend. When predicting founder success, the fraction of shortlisted founders who succeed increases from about 1 in 50 to 1 in 10, and to about 1 in 6 with a handful of expert-written questions. A second dataset which focuses on identifying failed loans, Kiva microloans, shows the same qualitative pattern. Across both tasks, the lesson is the same: a language model is most useful not as the predictor but as a generator of features a person can read, edit, and improve, and when positives are scarce that interpretability costs no measurable accuracy.
Large language models (LLMs) make next-token predictions based on clues present in their context, such as semantic descriptions and in-context examples. Yet, elucidating which prior tokens most strongly influence a given prediction remains challenging due to the proliferation of layers and attention heads in modern architectures. We propose Jacobian Scopes, a suite of gradient-based, token level causal attribution methods for interpreting LLM predictions. Grounded in perturbation theory and information geometry, Jacobian Scopes quantify how input tokens influence various aspects of a model's prediction, such as specific logits, the full predictive distribution, and model uncertainty (effective temperature). Through case studies spanning instruction understanding, translation, and in-context learning (ICL), we demonstrate how Jacobian Scopes reveal implicit political biases, uncover word- and phrase-level translation strategies, and shed light on recently debated mechanisms underlying in-context time-series forecasting. On token-ablation faithfulness benchmarks (LAMBADA and IWSLT2017) across six LLMs, the distribution-level Fisher and Temperature Scopes significantly outperform established gradient-based attribution methods in 10 of 12 model-dataset settings. To facilitate exploration of Jacobian Scopes on custom text, we open-source our implementations and provide a cloud-hosted interactive demo at https://huggingface.co/spaces/Typony/JacobianScopes.