WEBVTT

1
00:00:00.400 --> 00:00:04.679
An agent can fail, roll back, and repeat the same mistake.

2
00:00:04.799 --> 00:00:08.163
BeliefWeaver turns recovery into cumulative reasoning.

3
00:00:08.783 --> 00:00:15.476
It keeps a relational belief state: repair hypotheses, attempts, observations, and checkpoints.

4
00:00:15.596 --> 00:00:19.680
That evidence persists across attempts, including rollbacks.

5
00:00:20.300 --> 00:00:24.028
An implicit world model evaluates repair hypotheses.

6
00:00:24.148 --> 00:00:26.952
A diagnostic gate controls execution.

7
00:00:27.072 --> 00:00:31.531
New evidence updates belief, so the next attempt starts better informed.

8
00:00:32.151 --> 00:00:36.615
The paper distills four recurring roles into a 9B student.

9
00:00:36.735 --> 00:00:41.333
Supervised fine-tuning is followed by on-policy distillation.

10
00:00:41.953 --> 00:00:49.341
On 113 tasks: Terminus-2 reaches a 48.6% final score.

11
00:00:49.461 --> 00:00:55.269
BeliefWeaver reaches 53.7%; SFT + OPD reaches 58.4%.

12
00:00:55.889 --> 00:00:58.883
Recovery improves when evidence survives failure.

13
00:00:59.003 --> 00:01:02.002
Explore BeliefWeaver and read the paper.