diff --git "a/index.html" "b/index.html" --- "a/index.html" +++ "b/index.html" @@ -3,10 +3,10 @@ -Compressed Value Functions — can a memento replace the chain-of-thought? +Compressed Value Functions — inspector @@ -106,19 +101,17 @@
Compressed Value Functions

Can a short memento replace the chain-of-thought for value prediction?

A process value function estimates V = P(reach a correct answer | reasoning so far). - Reading a full reasoning trace is expensive. Each example below asks: does a compact memento of the - trace let the value model predict V as well as the full trace does? We show the ground-truth Monte-Carlo V - and the trained value-PRM's prediction V̂ read from the full trace and from five different mementos.

+ Pick a cut on the left; the pane compares the full reasoning trace against five mementos of it, side by + side, with the trained value-PRM's prediction V̂ read from each.

How to read this

-
-
true V — Monte-Carlo value: fraction of rollouts continuing from this partial - trace that reach a correct answer. The black tick on each bar marks it.
-
V̂ (prediction) — the trained MC-value PRM's - P(yes) when shown that representation. The colored bar is V̂; closer to the tick = better.
+
+
true V — Monte-Carlo value: fraction of rollouts from this partial trace that + reach a correct answer. The black tick on each meter marks it.
+
V̂ (prediction) — the trained MC-value PRM's P(yes) + when shown that representation. Bar = V̂; closer to the tick = better prediction.
Prediction accuracy | V̂ − V |
✓ close < 0.15 @@ -126,21 +119,25 @@ ✗ far ≥ 0.35
-
+

Aggregate zero-shot value prediction

-
+
-

Examples

-
+

Inspect a cut

+
+
+
+
-
Generated from value_showcase.json · trained value-PRM P(yes) on held-out - MC-val cuts · sources scored on the same cut. Self-contained page — data embedded.
+
Trained value-PRM P(yes) on held-out MC-val cuts; every representation scored on the same + cut. Self-contained — data embedded; LaTeX via KaTeX. ← / → to page through cuts.