Papers
Latest Papers
83 resultsVision-language-action policies increasingly use a deterministic action head trained with an $\ell_1$ regression loss, and recent failure-detection work treats that head as carrying no predictive distribution, adding ensembles, labeled rollouts, or learned recalibrators to obtain a reliability signal. We observe that a…
At a fixed completion-token budget, a model can spend its tokens on its own chain of thought or on exact tool round-trips paid for out of the same budget, and harnesses set the two caps without knowing the rate at which one converts into the other. Test-time-compute research prices model-side options against each other…
Given only anonymous population length summaries, categorical pairs narrow the uncertainty in a fixed attention backend's expected padding work relative to marginals, but do not always settle which backend to use. We characterize that information tradeoff for padded work equal to batch size times maximum length and spl…
A recovery query can reveal whether a state-changing operation succeeded while consuming the service that could finish it. We compare recovery policies in a synthetic known-parameter model where execution, status acquisition and receipt delivery share service and pending attempts retain their late effects. Across a fro…
Memory-efficient training increasingly drops the fp32 master copy of the weights and rounds every update into a low-precision format, yet its loss cost is known only through upper bounds. We compute it as an exact moment identity. For SGD on a quadratic whose iterate is stored by stochastic rounding, the stationary cov…
Fixed XOR masks in an archived reproducible sampler induce a rank-associated winner law that persists over a sufficient unequal-logit neighborhood. Independent Gumbel-max correctness and uniform hash marginals do not specify this shared-word law. We derive the masks' binary-trie selection probabilities and a rational p…
Input-embedding rows receive a gradient only when their token occurs, and Adam(W)'s effect on such a row is known only through a one-step bound and an unsaturated frequency scaling. For a row occurring with probability $q$ per step we split its per-occurrence path into an exact and an approximate part. The exact part i…
On the persona prompts of two published patient simulators, PatientSim and AgentClinic, played by Ministral-3-14B, we measure what an LLM history-taker could gain by choosing how to ask from the patient's persona: the headroom of a persona-contingent manner over the best single manner, which bounds what acting on the p…
Cooperative teams of frozen large language models can adapt without gradients by storing and reusing their own experience, but a team paid only a terminal win-or-lose signal is given no step-level credit, and manufacturing one without an evaluator or gradients has never been measured. We introduce CARE (Contrastive Adv…
Test-time sampling budgets carry two components. How many chain-of-thought samples to draw has an industry of stopping rules; the token cap that bounds each sample is inherited from the serving stack, taken as given by the studies that size budgets under it, and almost never audited against the models it bounds. We sep…
Raising a MedMNIST+ classifier's input from 64 to 128 px roughly quadruples its compute, which could instead buy ensemble members; resolution benchmarks do not hold compute fixed. At the compute of one forward pass of a 128 px ResNet-18 we compare two concrete ways to spend it: three 64 px SmallCNNs, averaged and tempe…
Generated 3D scenes are increasingly judged by a settle test: simulate the scene for a few seconds and call it plausible if nothing moved past a threshold. Each paper fixes one engine, step, window and threshold, and none reports how much the verdict depends on that choice; the simulator-free alternative, a rigid-block…
Autoregressive diffusion emulators are validated on one-step errors at held-out states but used for long rollouts. We ask how much of an emulator's long-run bias in an observable is determined by those one-step errors, weighted by the true dynamics' response. A classical perturbation identity, evaluated at the true sta…
Curricula and self-play proposers in RL with verifiable rewards pick tasks by pass rate, reading the variance of success as what a task teaches. Under a stochastic verifier such as a sampled LLM judge, that variance sums a learnable signal, the spread of the verifier's mean across responses, and per-call coin noise. We…
First-failure feedback can reveal information about a reused hidden test suite, but an optimal information advantage need not be available to a finite repair generator or realized by a public-data selector. We separate these quantities in a two-attempt validation experiment with fixed initial bugs, syntax-generated can…