AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

Bits That Count: Quantifying and Predicting Capabilities of Language Models

Elizabeth Donoway, Hailey Joren, Michael R DeWeese, Ethan Perez, John Schulman, Fabien Roger, Jan Leike

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Gradient-free adaptation and merging

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
pre-teach multiplication with operator notation before fine-tuning (ablation)TinyStories-1B LoRA · Multiplicationcapacity threshold EDL/P*: 1.02 -> 0.05 bits/parameterImproved
pre-teach a different operation (addition/subtraction) before multiplication (ablation)TinyStories-1B LoRA · Multiplicationcapacity threshold EDL/P*: 1.56 bits/parameter, still increasing returnsNo clear effect
pre-elicit format with a single related example (ablation)Llama 3.2 1B · Addition/subtractionEDL per token: reduces EDL/D by approximately an order of magnitudeImproved
fine-tune on a single random example instead of using the instruct modelLlama 3.1 8B base · Addition/subtractionaccuracy: 0% -> 96% (Instruct 56.1%)Improved
scale model size (ablation)Llama 3 1B/3B/8B · Multiplication, addition/subtractioncapacity threshold EDL/P*: larger models have lower thresholdsImproved
use LoRA below capacity threshold vs above (ablation)LoRA · Elicitation taskscapacity vs full fine-tuning: ~0.01-0.1 bits/parameter for elicitation vs ~1 for teachingHurt
distill R1 reasoning traces by SFTQwen2.5 models · multi-domain reasoningaccuracy: up to 23 pp from single exampleImproved
replace EDL with plain MDL (ablation)capability regime analysis · elicitation vs teaching classificationdiscriminationHurt
scale dataset to millions of examplesLlama models · MultiplicationEDL signature: elicitation-like to teaching-like crossoverMixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy