Track genomics, neuroscience and molecular biology.
Search papers, labs, and topics across Lattice.
This study investigates the source of performance gains in adaptive computation for language models by comparing layer-skipping and repetition programs against input-blind controls. Using 32 programs across two models and 4,413 multiple-choice items, the research demonstrates that input-blind perturbations can yield substantial oracle headroom, often exceeding that of the real programs, particularly when option order is consistent. The findings suggest that a significant portion of the observed gains may stem from factors other than the specific adaptive computation strategy, such as fixed offsets or sensitivity to option ordering, rather than a unique benefit of the selected layer computation itself.
Input-blind controls can account for substantial oracle headroom in adaptive computation, challenging the assumption that observed gains directly reflect the efficacy of selected layer programs.
Adaptive computation aims to improve language-model inference by tailoring execution to each input. For layer programs, oracle evaluations use known answers to estimate the potential gain from this flexibility, before a practical selector is available. However, a gain from selection does not by itself explain why the chosen programs help. This study examines this distinction using 32 layer-skipping and repetition programs on two models and 4,413 multiple-choice items. The analysis compares their gains over a fixed action selected without the evaluation prompt with those of input-blind perturbations at the same sites, re-evaluating selections on another prompt. With shared option order, the controls give 10.2-11.8 and 15.6-19.4 percentage points of headroom on Qwen3-4B-Base and Llama-3.1-8B, exceeding the real programs' 9.0 and 10.1 in all three random-direction draws per model. They match answer-change rate only, and the ordering depends on the menu: in post hoc comparisons, real programs lead on Llama's repeat-only menu in every draw. A smaller KL-calibrated comparison, including an input-dependent control, favours real programs in point estimate, with inconclusive corrected tests. Fixed letter offsets produce headroom of similar scale. Rotating options sharply reduces both families' headroom, while leaving positive real-minus-control differences of 1.4-2.3 and 3.7-4.5 points; their magnitudes and statistical support depend on further adjustments and the reference. A supplementary generated-answer test finds that search-selected programs keep a 26.0-point advantage over programs selected for other problems after rewording, without a placebo comparison. These results show that substantial headroom can persist across prompts with shared option order without establishing a benefit specific to the selected layer computation; neither ordering against these controls identifies that benefit.