wassname/moral-maps
1
0
Fork 0
mirror of https://github.com/wassname/moral-maps.git synced 2026-09-09 11:27:22 +08:00
Code Issues Packages Projects Releases Wiki Activity

README: add think-budget sensitivity (UAT2 -- steer delta grows with think)

Browse Source 
MFQ-2 mean |steer delta| rises monotonically with the think budget: 0.068 (1) ->
0.149 (64) -> 0.319 (128) -> 0.682 (256), pmass >= 0.95. Proves the unified reader's
think budget carries the steer. Past ~512 the model closes </think> and the
readout collapses (coherent ceiling). Source: ablation_think_budget.py, job 228.

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
This commit is contained in:
wassname
2026-06-25 03:31:15 +08:00
co-authored by Claudypoo
parent d2a5cab4c9
commit 1c8326d0ea
1 changed files with 7 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+7
View File
@@ -195,6 +195,13 @@ violations across foundations (Care +0.65 nats) while lowering Social Norms (the
-0.24); `-C` does the opposite (Care -0.31, Social Norms +1.52). Both poles stay coherent. The MFV
figure below shows the two arms moving apart.
The readout earns the steer through its think budget: the survey reader generates the think tokens,
the activation steer accrues over them, then the answer slot is read. So the same vector moves the
profile more when given more think. On MFQ-2 the mean per-foundation `|steer delta|` grows with the
budget, 0.068 (1 think token) -> 0.149 (64) -> 0.319 (128) -> 0.682 (256), with `pmass` staying >= 0.95.
(Past ~512 the model closes `</think>` on its own and the readout collapses, so the budget has a
coherent ceiling.)
## One vector across every instrument
The same calibrated Authority/Care vector, administered through every instrument tinymfv supports