AC
⭐ GitHub
EACL 2026 Findings · Published 04 Jan 2026

Debiasing LLMs
with Adaptive Causal
Prompting & SoT

ACPS routes each query to standard or conditional front-door adjustment and uses Sketch-of-Thought mediators to cut token costs while keeping causal faithfulness and robustness across reasoning tasks.

Read the Paper View Code Watch Demo
Related papers FAIR-ToT CA2KG
7

Benchmarks

↓58%

Token Reduction

2×

Routing Methods

✓

Robust

ACPS Causal Graph

Front-door & Conditional Front-door

ACPS causal diagram showing front-door and conditional front-door adjustment

Instance-level causal routing plus Sketch-of-Thought mediators delivers robust, efficient reasoning across heterogeneous tasks.

Methodology

Adaptive Causal Prompting + Sketch-of-Thought

A novel framework combining instance-level causal routing with efficient reasoning mediators

Instance-level Routing

Automatically choose standard or conditional front-door adjustment based on query context to preserve causal faithfulness.

Sketch-of-Thought Mediators

Diverse sampling and clustering produce compact mediators that trim token usage without losing reasoning coverage.

Efficiency & Robustness

Tested on GSM8K, HotpotQA, FEVER, Math, StrategyQA, MusiQue, CommonsenseQA; includes injected/shuffled stress tests.

Experiments

Benchmarks & Notebooks

Core ACPS Experiments

  • CommonsenseQA — acps-on-commonsenseqa.ipynb
  • FEVER — acps-on-fever.ipynb
  • HotpotQA — acps-on-hotpotqa.ipynb
  • GSM8K — acps-on-gsm8k.ipynb
  • Math — acps-on-math.ipynb
  • StrategyQA — acps-on-strategyqa.ipynb
  • MusiQue — acps-on-musique.ipynb

Datasets & Studies

  • SoT builders — sots_datasets/*.ipynb
  • Efficiency comparison — efficiency_comparison/
  • Robustness (shuffled/injected) — robustness_study/
  • Helpers (encoders, prompts) — helpers/
Browse all notebooks on GitHub →

Where ACPS Helps

Robust, token-efficient reasoning across heterogeneous tasks

🧮

Math & GSM8K

Stable reasoning with fewer tokens for complex mathematical problems.

📚

Multi-hop QA

HotpotQA, StrategyQA, MusiQue with adaptive causal routing.

🧾

Fact Verification

FEVER robustness under injected/shuffled contexts.

🧠

Commonsense

CommonsenseQA causal mediation with Sketch-of-Thought.

Demo Video

ACPS demo

Why ACPS?

Adaptive causal routing plus Sketch-of-Thought mediators keeps reasoning faithful while reducing token usage significantly.

Paper at a Glance

  • Venue: EACL 2026 Findings
  • Published: 04 Jan 2026
  • License: MIT (code), CC BY 4.0 (paper)

Ready to Reproduce ACPS?

Open the notebooks, regenerate Sketch-of-Thought mediators, and run the causal routing pipeline on your own benchmarks.

Open README View Notebooks