Reproduce “Know your audience: specializing grounded language models with listener subtraction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Aaditya K. Singh, David Ding, Andrew Saxe, Felix Hill, Andrew K. Lampinen
The authors train a grounded captioner through a two-listener reference game, rewarding captions that improve one frozen listener’s target identification over another’s. A learned adapter modifies a frozen CLIP-to-language model using REINFORCE. With COCO and a grayscale listener, listener-accuracy difference reaches 0.156, versus 0.019 for the strongest prompt baseline. This demonstrates specialization to imposed perceptual differences, not human personalization.
Key insights
- The contrasting presence of multiple listeners with differing perceptual capabilities is essential for the emergence of targeted, adaptive language specialization.
- A contrastive reward signal, which incorporates distractor images, is crucial for encouraging specialization while simultaneously preserving the relevance and grounding of the generated language.
- Complex, context-dependent language specialization can arise solely from interaction and reward signals, bypassing the need for explicit linguistic supervision or extensive prompt engineering.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.