Contrastive negatives and temperature define a comparison task, not automatic semantic truth

Suppose both encoders produce nonzero features projected into a shared dimension, then normalize each feature to unit length. Their dot product is cosine similarity under that representation. Divide similarities by a positive temperature to form logits, or multiply by an explicitly defined inverse-temperature scale. These conventions are reciprocals; confusing them reverses a statement about making a softmax sharper. The original CLIP implementation in its paper learns a log-parameterized multiplicative scale, so describing that parameter requires distinguishing it from a divisor temperature.

For one aligned image row, the loss is negative log probability of its intended text among the batch candidates. A common symmetric objective averages image-to-text and text-to-image losses. This depends on the gathered comparison set and target indices, not merely the batch size. A larger batch changes the denominator, possible hard examples, duplicate frequency, and compute. It does not make every additional candidate a valid semantic negative. Keep normalization, target identity, gather order, and the scale parameter visible when diagnosing a changed loss.

With one positive target, increasing the score of any other candidate increases its contribution to the denominator, even if that candidate is a valid paraphrase or another view of the same content. That is a false-negative issue relative to the intended semantic task, not a tensor error. Asset duplicates, repeated captions, closely related frames, and broad descriptions all create such cases. Grouped batching, duplicate-aware masking, or multi-positive objectives can address particular mechanisms, but each changes the learning problem and needs clearly defined positive relations.

A lower temperature emphasizes score differences and often concentrates probability on the highest-scoring candidates. It can also make a mislabeled or false-negative comparison more consequential. It does not improve the actual encoder representations by itself at evaluation time, and it does not make a row softmax a calibrated probability that an image-text pair is universally true. Probability depends on the offered candidate set; adding another compatible caption can redistribute mass even when the underlying pair's features are unchanged. Evaluate retrieval and semantic tasks separately from training-set discrimination.

Trace positive IDs and inspect the highest-scoring alleged negatives. Measure duplicates and semantic overlap by source and cohort. Compare a controlled change in batch construction or temperature while preserving the declared evaluation task, and check gradient finiteness with stable log-space computation. If multi-positive training is adopted, test its label normalization and retrieval behavior rather than reporting its loss value as directly comparable with a changed one-positive denominator. A fluent description or one successful retrieval is also insufficient to establish relational grounding across unseen combinations.

The vector lab illustrates cosine with a fixed unit reference along the positive x-axis. Reset, set both angle 0° and magnitude 1, predict x-projection 1, then inspect or Play. Reset and set both angle 60° and magnitude 1 for x-projection 0.5. Both selected vectors are unit length, so these x-components equal their dot products with that reference. The instrument contains one two-dimensional vector and no encoder, comparison batch, negative labels, temperature, softmax, or gradient. Use it only to verify the geometry entering a score, not to simulate contrastive learning.