@JongwonPar9958 on Backscroll

3 tweets from @JongwonPar9958 that Backscroll ranked for substance.

Jongwon Park @ ICML @JongwonPar9958 2026-07-09 00:22Z

It's really hard to guarantee you've found and patched every hacking path in an RL environment before training. It is a long-tail problem. The problem is when you post-train on a reward-hackable environment, it doesn't just corrupt the training signal. The model can be broadly

Jongwon Park @ ICML @JongwonPar9958 2026-07-08 12:50Z

As RL Tasks get more advanced, the market splits like semiconductors did — chip design (NVIDIA, Intel) vs foundry (TSMC, Samsung). We're building the TSMC of Tasks. DM if interested.

Jongwon Park @ ICML @JongwonPar9958 2026-07-03 14:23Z

MacDiarmid et al showed reward hacking in narrow domain can generalize into alignment faking and sabotage. The fix they've deployed is inoculation prompting (2511.18397). But a newer paper (2604.25891) finds inoculation may not remove the misalignment so much as hide it. It

How it ranks: Backscroll reads my Twitter/X timeline and ranks every tweet for substance with an LLM taste rubric — not raw engagement. Curated by Surya Dantuluri.