From a3e30aceaa5ee7435dcd47762506372238bf59b6 Mon Sep 17 00:00:00 2001
From: wassname <1103714+wassname@users.noreply.github.com>
Date: Fri, 21 Aug 2026 13:45:14 +0800
Subject: [PATCH] Add the DailyDilemmas results table and the CoLoRAI
 acceptance to the README

The headline numbers only lived in the arXiv PDF, so anything scraping the
repo could not see them.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
---
 README.md | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/README.md b/README.md
index 11e2007..23a783e 100644
--- a/README.md
+++ b/README.md
@@ -9,6 +9,8 @@
 
 *Serving up data-efficient inner alignment, one satisfying rotation at a time.*
 
+Accepted to the CoLoRAI workshop, ICML 2026.
+
 > Gradient-based honesty steering trained as an adapter on the model's own representations, not outputs. Human input: two contrasting words, no preference labels.
 
 **How it works:** Train a single adapter (~1 hour on Gemma-3-1B). At inference, dial the steering coefficient: +1 for more honest, -1 for less, 0 for baseline. One adapter, bidirectional control.
@@ -23,6 +25,17 @@ Applications:
 
 ![Bidirectional control](docs/img/fig_bidirectional_demo.svg)
 
+## Results
+
+Train on 800 honesty persona pairs, test on [DailyDilemmas](https://arxiv.org/abs/2410.02683), an external benchmark of 1,360 moral dilemmas across 9 value dimensions, built independently of this work. gemma-3-1b-it, n=3 seeds.
+
+| Method | Steer F1 | Flip% | Arb | Focus | Coh |
+|:-------|---------:|------:|----:|------:|----:|
+| AntiPaSTO | 14.1±3.3 | 13% | 0% | 10.5 | 0.6 |
+| Prompting | 4.5 | 12.7% | 0% | — | -0.09 |
+
+3.1x the Steering F1 of prompting, with fewer side effects. Full tables, ablations, and cross-model results in the [paper](https://arxiv.org/abs/2601.07473).
+
 ## Quick Start
 
 ### Bake your own