Contrastive Language Models

Commenters focused on the Contrastive Language Models idea: taking a frozen LLM, adding a small model that reads hidden states plus a question to produce answer probabilities, enabling non-autoregressive, fast decision-making. amluto praised the approach but argued it’s fundamentally different from classification versus agentic, sequential decision problems and wondered how “reasoning” fits into the framework. Others were enthusiastic: andy12 likened it to “CLIP for actions,” rahimnathwani traced roots to robotics work and linked the new repo, and mugul welcomed the experimental transparency about what did and didn’t work. Several noted the appeal of decoupling input and output token spaces for speed and cost, and some proposed variants (e.g., asking the model whether further thinking is needed).

Opinion split on evaluation, latency, and terminology. Some users reported similar runtimes to Jev on local hardware (sdan, brookman64k) while eadwu and mugul questioned the relevance of latency when network hops dominate; r0x0r007 and 0x4139 flagged imperfections on game benchmarks and limits on typed-decision tasks like arithmetic or negation chains. fxwin objected to calling the method “System One” as a buzzword, while tancop and TeMPOraL defended the instinctive analogy. in-silico and others were curious about timing relative to Jev’s release. Overall, the divide centered on how meaningful the latency claims and benchmarks are and whether the contrastive approach generalizes to true agentic reasoning.