Rethinking Few Shot CLIP Benchmarks: A Critical Analysis in the Inductive Setting

CLIP is a foundational model with transferable classification performance in the few-shot setting. Several methods have shown improved performance of CLIP using few-shot examples. However, so far, all these techniques have been benchmarked using standard few-shot datasets. We argue that this mode of evaluation does not provide a true indication of the inductive generalization ability using few-shot examples. As most datasets have been seen by the CLIP model, the resultant setting can be termed as partially transductive. To solve this, we propose a pipeline that uses an unlearning technique to obtain true inductive baselines. In this new inductive setting, the methods show a significant drop in performance (-55% on average among 13 baselines with multiple datasets). We validate the unlearning technique using oracle baselines. An improved few-shot classification technique is proposed that consistently obtains state-of-the-art performance over 13 other recent baseline methods on a comprehensive analysis with 5880 experiments - varying the datasets, differing number of few-shot examples, unlearning setting, and with different seeds. Thus, we identify the issue with the evaluation of CLIP-based few-shot classification, provide a solution using unlearning, propose new benchmarks, and provide an improved method.

Addressing Blind Guessing: Calibration of Selection Bias in Multiple-Choice Question Answering by Video Language Models

Evaluating Video Language Models (VLMs) is a challenging task. Due to its transparency, Multiple-Choice Question Answering (MCQA) is widely used to measure the performance of these models through accuracy. However, existing MCQA benchmarks fail to capture the full reasoning capabilities of VLMs due to selection bias, when models disproportionately favor certain answer options based on positional patterns observed during training. In this work, we conduct a comprehensive empirical analysis of several VLM architectures across major datasets designed to assess complex video-focused reasoning. We identify where the bias is most pronounced and demonstrate to what extent model responses reflect genuine understanding of video content and related questions, as opposed to reliance on arbitrary patterns or superficial cues, such as answer position. By decomposing the MCQA task and adapting fairness bias metrics to VLMs, we introduce a post-processing calibration technique BOLD to balance this bias. Our results show that reducing selection bias improves not only debiasing metrics but also overall model performance, including Accuracy and F1 Mean score. Our method, by suppressing "blind guessing", offers a more cost- and time-effective approach to mitigating selection bias compared to existing techniques. This study represents the first focused investigation of selection bias in video-to-text LLM-powered models.

CLIP Adaptation by Intra-modal Overlap Reduction

Numerous methods have been proposed to adapt a pre-trained foundational CLIP model for few-shot classification. As CLIP is trained on a large corpus, it generalises well through adaptation to few-shot classification. In this work, we analyse the intra-modal overlap in image space in terms of embedding representation. Our analysis shows that, due to contrastive learning, embeddings from CLIP model exhibit high cosine similarity distribution overlap in the image space between paired and unpaired examples affecting the performance of few-shot training-free classification methods which rely on similarity in the image space for their predictions. To tackle intra-modal overlap we propose to train a lightweight adapter on a generic set of samples from the Google Open Images dataset demonstrating that this improves accuracy for few-shot training-free classification. We validate our contribution through extensive empirical analysis and demonstrate that reducing the intra-modal overlap leads to a) improved performance on a number of standard datasets, b) increased robustness to distribution shift and c) higher feature variance rendering the features more discriminative for downstream tasks.