paper

On the Limits of Prompt-Conditioned Language Models as General-Purpose Learners

paperactiveprovisionalon-the-limits-of-prompt-conditioned-language-models-as-general-purpose-learners-08c7d3f3·1 events·first seen 5h ago

Aliases: On the Limits of Prompt-Conditioned Language Models as General-Purpose Learners

Co-occurring entities

PAC-Bayes

More like this (12)

Concept-Constrained Prompt Learning Language Models are Few-Shot Learners Reasoning Language Models Unintended Effects of Geographic Conditioning in Large Language Models Latent Context Language Models Reinforcement Learning for Language Models Dango: A Strictly L1-Only Large Language Model for Studying Second Language Acquisition When English Isn't the Best Teacher: Source Language Effects in Cross-Lingual In-Context Learning Reinforcement Learning Elicits Contextual Learning of Unseen Language Translation multi-turn language models Arithmetic Pedagogy for Language Models Language Generation in the Limit

Recent events (1)

6arXiv · cs.LG·5h ago·source ↗

PAC-Bayes analysis establishes formal expressivity and alignment floors for prompt-conditioned LLMs

A new arXiv preprint models user-LLM interaction as a bilevel cheap-talk game and derives PAC-Bayes bounds showing two irreducible limitations: an 'expressivity floor' where language's finite channel capacity makes distinct tasks indistinguishable, and an 'objective-misalignment floor' where alignment constraints prevent reaching user-ideal outputs. The authors prove that prompt-conditioned LLMs cannot be universal problem solvers, as correct behavior on certain task families is provably unattainable even with infinite data, optimal training, or model scaling. The work suggests multimodal inputs and external memory as potential mitigations by increasing task-relevant information bandwidth.

Evaluation and Benchmarking Alignment and RLHF PAC-Bayes On the Limits of Prompt-Conditioned Language Models as General-Purpose Learners