Works · Live Demo

ACPS ·
Adaptive Causal Prompting with Sketch-of-Thought

A framework for debiasing black-box LLMs: instead of trusting a single chain of reasoning, ACPS samples several reasoning paths, estimates how causally load-bearing each one is, then lets the paths vote on the final answer — weighted by their estimated causal effect. Sketch-of-Thought mediators keep the reasoning compact, cutting token usage by ~58% without hurting accuracy.

EACL 2026 Findings 7 benchmarks · GSM8K, HotpotQA, FEVER, Math, CommonsenseQA Open full app ↗

How it works

Four stages, one causal-weighted answer

01

Generate reasoning paths

The query is routed adaptively and expanded into several independent Sketch-of-Thought reasoning paths toward an answer.

02

Estimate causal effect

Each path is treated as a mediator; front-door (or conditional front-door) adjustment estimates how much it causally drives the answer.

03

Causal-weighted vote

Candidate answers are aggregated, weighting each path's vote by its estimated causal effect rather than counting paths equally.

04

Final answer

The highest-weighted answer is returned, along with the reasoning paths that supported it — debiased against spurious shortcuts.

Try it

Live playground

Open in new tab ↗

This is the real ACPS pipeline, hosted on Cloudflare Workers — enter a question or use the default prompt to see reasoning paths generated, scored, and voted on live.

If the embedded app doesn't load in your browser, open it directly ↗.