9 problems
- 0 votes0 replies0 views
Effective Lipschitz concentration for transformer language models
Effective Lipschitz concentration. The effective Lipschitz constant satisfies , specifically , whil…
- 0 votes0 replies1 view
The reward-hacking conjecture for Xent Games
A verifiable task is one in which a judge model can enforce the legality of player moves and score their plausibility, thereby providing a continuous reward signal for search optim…
- 0 votes0 replies2 views
The relaxed Hilberg hypothesis for block entropy
Let be a stationary stochastic process on a finite alphabet . Write for its block entropy and let be its Shan…
- 0 votes0 replies0 views
Applicability of sparse Bregman projection beyond language-model decoding
A probability vector is sparsified by solving an -regularized separable Bregman projection problem over the simplex; the resulting primitive admits exact and efficient opti…
- 0 votes0 replies0 views
The entropy concavity conjecture for language-model phase space
Let be the phase space associated with a language-model geometrization, and let … be a differentiable extension of the language-model entropy function. The entro…
- 0 votes0 replies0 views
The smooth-extension conjecture for thermodynamic functions of language models
Let be the vocabulary, let be the finite set of sentences of length , and let a geometrization consist of an embedding…
- 0 votes0 replies0 views
The AGI finite-determination conjecture for language-model entropy
Let be the set of sentences, let be a causal language model, and let … be its entropy function. The AGI finite-determination conjecture. If a causal lan…
- 0 votes0 replies0 views
The distributional embedding conjecture for intelligent language models
Let be the set of sentences, let be a causal language model, and let … map each sentence to its output-probability distribution, where…
- 0 votes0 replies1 view
Conjecture that language-model key-value-query structure is approximated by a three-layer linear network
A language model has a key-value-query structure associated with its attention mechanism. Key-value-query approximation conjecture. The key-value-query structure in language models…