---
format: "aidr-story-markdown/v1"
id: "c42330bc780a526148da04cd54d03f779fb331d39e566c30676f4fb8a6886d94"
canonical_url: "https://aidr.today/c42330bc?lang=en"
title: "Theoretical Limits of Language Model Alignment"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-11-12T00:00:00.000Z"
category: "Research"
topics: ["llm"]
source_urls: ["https://machinelearning.apple.com/research/theoretical-limits-alignment"]
summary: "A study titled Theoretical Limits of Language Model Alignment derives the maximum achievable expected reward gain for a fixed KL-divergence budget using a Jeffreys divergence term, and shows that reward ensembling mitigates reward hacking in proxy reward settings."
---

# Theoretical Limits of Language Model Alignment

> [Open the canonical story](<https://aidr.today/c42330bc?lang=en>)

**Published:** 2026-11-12T00:00:00.000Z
**Category:** Research
**Topics:** llm

## Summary

A study titled Theoretical Limits of Language Model Alignment derives the maximum achievable expected reward gain for a fixed KL\-divergence budget using a Jeffreys divergence term, and shows that reward ensembling mitigates reward hacking in proxy reward settings\.

## Sources

- [Story source](<https://machinelearning.apple.com/research/theoretical-limits-alignment>)