Retitle: Weight-Space Geometry of Offline Reasoning Training
Browse files- README.md +4 -4
- data/meta.json +2 -2
- index.html +1 -1
README.md
CHANGED
|
@@ -1,5 +1,5 @@
|
|
| 1 |
---
|
| 2 |
-
title:
|
| 3 |
emoji: 🧭
|
| 4 |
colorFrom: indigo
|
| 5 |
colorTo: pink
|
|
@@ -10,10 +10,10 @@ license: mit
|
|
| 10 |
short_description: Interactive weight-space geometry of six reasoning losses
|
| 11 |
---
|
| 12 |
|
| 13 |
-
#
|
| 14 |
|
| 15 |
-
An **interactive blog** companion to the ICML 2026 MechInterp Workshop paper
|
| 16 |
-
|
| 17 |
*identical* math rollouts from one base model (Qwen3-4B, attention-only LoRA). The page lets
|
| 18 |
you turn the same knobs the authors did — pick method pairs, scrub across all 36 transformer
|
| 19 |
layers, and watch the geometry respond.
|
|
|
|
| 1 |
---
|
| 2 |
+
title: Weight-Space Geometry of Offline Reasoning Training
|
| 3 |
emoji: 🧭
|
| 4 |
colorFrom: indigo
|
| 5 |
colorTo: pink
|
|
|
|
| 10 |
short_description: Interactive weight-space geometry of six reasoning losses
|
| 11 |
---
|
| 12 |
|
| 13 |
+
# Weight-Space Geometry of Offline Reasoning Training
|
| 14 |
|
| 15 |
+
An **interactive blog** companion to the ICML 2026 MechInterp Workshop paper
|
| 16 |
+
(*"Same Data, Different Losses, Same Circuits?"*). Six offline reasoning losses (SFT, RFT, DFT, RIFT, Offline GRPO, DPO) are trained on
|
| 17 |
*identical* math rollouts from one base model (Qwen3-4B, attention-only LoRA). The page lets
|
| 18 |
you turn the same knobs the authors did — pick method pairs, scrub across all 36 transformer
|
| 19 |
layers, and watch the geometry respond.
|
data/meta.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
{
|
| 2 |
-
"title": "
|
| 3 |
-
"subtitle": "
|
| 4 |
"venue": "ICML 2026 · Mechanistic Interpretability Workshop",
|
| 5 |
"repo": "https://github.com/AlexWortega/capabilityvectors",
|
| 6 |
"methodColors": {
|
|
|
|
| 1 |
{
|
| 2 |
+
"title": "Weight-Space Geometry of Offline Reasoning Training",
|
| 3 |
+
"subtitle": "Same data, different losses, same circuits? Six offline reasoning losses, one base model, identical math rollouts — do they converge in weight space?",
|
| 4 |
"venue": "ICML 2026 · Mechanistic Interpretability Workshop",
|
| 5 |
"repo": "https://github.com/AlexWortega/capabilityvectors",
|
| 6 |
"methodColors": {
|
index.html
CHANGED
|
@@ -3,7 +3,7 @@
|
|
| 3 |
<head>
|
| 4 |
<meta charset="utf-8" />
|
| 5 |
<meta name="viewport" content="width=device-width, initial-scale=1" />
|
| 6 |
-
<title>
|
| 7 |
<meta name="description" content="An interactive look at the weight-space geometry of six offline reasoning losses trained on identical data." />
|
| 8 |
<link rel="preconnect" href="https://fonts.googleapis.com" />
|
| 9 |
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin />
|
|
|
|
| 3 |
<head>
|
| 4 |
<meta charset="utf-8" />
|
| 5 |
<meta name="viewport" content="width=device-width, initial-scale=1" />
|
| 6 |
+
<title>Weight-Space Geometry of Offline Reasoning Training</title>
|
| 7 |
<meta name="description" content="An interactive look at the weight-space geometry of six offline reasoning losses trained on identical data." />
|
| 8 |
<link rel="preconnect" href="https://fonts.googleapis.com" />
|
| 9 |
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin />
|