New paper uses RL post-training to push LLMs toward counterfactual-invariant fairness

nagpalchirag · x · 2026-10-27

• A new paper, Aligning Language Models towards Counterfactual Invariance, shows RL post-training improves reasoning and instruction following, but these gains may be disparate across prompt categories, user subgroups, and sensitive attributes.
• The team built on existing open-source alignment tooling (OSS reward models + GRPO) to derive a method letting LLMs use demographic data with more discretion and counterfactual-invariant fairness.
• Second paper from intern Basaar Demir, advised by @nagpalchirag.

Original post →

More from Safety

• Abliterated LLM finds $3,300 in live ElevenLabs keys; OpenAI/Anthropic keys already drained — deedydas · 2026-10-27
• Tailscale confirms serious Tailworm SSH flaws; fix due within 24 hours — AIFlow_ML · 2026-10-27
• OpenAI details how it locked down frontier training environments and red-teamed them with its own agents — AxSaucedo · 2026-10-27
• GMO hack: account compromises real, ransomware encryption still unconfirmed — jd_pressman · 2026-10-27
• Inference.net serves refusal-removed DeepSeek V4.1 Flash Cyber at cost, KYC required — AccBalanced · 2026-10-27
• Reddit asks: why can AI firms censor models so strictly yet fail to stop jailbreaks? — mikehocalate · 2026-10-27

Safety channel →