Yuhan Luo(@_yuhanluo):i’ve often worried that pass@k = 0 tasks weren’t worth shipping, but this makes me feel better about making extremely hard tasks -- it seems like failed traces can still yield useful learning signals, even when there’s no good solution to copy also this reminds me of a similar attempt: https://t.co/kL3Q8hUjdi RLTL;DR: failed attempt + verifier feedback -> generate TL;DR lesson -> train the "task to lesson" mapping (no separate teacher) HINT-SD: full failed attempt -- thru hindsight analysis --> select relevant actions + generate corrective feedback -> train that "trace to action" mapping at selected steps, with hindsight informed teacher predictions (the teacher that comes from the student)

2026.10.04 04:28

i’ve often worried that pass@k = 0 tasks weren’t worth shipping, but this makes me feel better about making extremely hard tasks -- it seems like failed traces can still yield useful learning signals, even when there’s no good solution to copy also this reminds me of a similar attempt: RLTL;DR: failed attempt + verifier feedback -> generate TL;DR lesson -> train the "task to lesson" mapping (no separate teacher) HINT-SD: full failed attempt -- thru hindsight analysis --> select relevant actions + generate corrective feedback -> train that "trace to action" mapping at selected steps, with hindsight informed teacher predictions (the teacher that comes from the student)

显示更多

Michael Kirchhof@mkirchhof_

2026.09.30 13:42

We let an agent self-improve by exploring and internalizing ”on this sort of task, keep this sort of thing in mind“ bits of self-feedback. It learned to solve tasks where the original policy failed 128 attempts in a row and GRPO flatlined. Paper: 🧵1/8

显示更多