# Revisiting the Superficial Alignment Hypothesis

arXiv: https://arxiv.org/abs/2410.03717 · HTML: https://mlsanity.com/paper/2410.03717/

## Key facts

- Cited by 12 papers in the map (data through 14 Oct 2026).
- Placed in LLM preference optimization methods, one of 10 directions in LLM alignment / RLHF / preference optimisation.
- Builds on 30 of its 35 references that are in the map.
- First posted on arXiv 27 Sep 2024.

## Placed in

- LLM alignment / RLHF / preference optimisation / [LLM preference optimization methods](https://mlsanity.com/alignment/llm-preference-optimization-methods/)

## Builds on (most cited first, 30 of 30)

- [Attention Is All You Need](https://mlsanity.com/paper/1706.03762/) (2017)
- [Language Models are Few-Shot Learners](https://mlsanity.com/paper/2005.14165/) (2020)
- [GPT-4 Technical Report](https://mlsanity.com/paper/2303.08774/) (2023)
- [Training language models to follow instructions with human feedback](https://mlsanity.com/paper/2203.02155/) (2022)
- [The Llama 3 Herd of Models](https://mlsanity.com/paper/2407.21783/) (2024)
- [Llama 2: Open Foundation and Fine-Tuned Chat Models](https://mlsanity.com/paper/2307.09288/) (2023)
- [Training Verifiers to Solve Math Word Problems](https://mlsanity.com/paper/2110.14168/) (2021)
- [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://mlsanity.com/paper/2005.11401/) (2020)
- [Scaling Laws for Neural Language Models](https://mlsanity.com/paper/2001.08361/) (2020)
- [Finetuned Language Models Are Zero-Shot Learners](https://mlsanity.com/paper/2109.01652/) (2021)
- [Mistral 7B](https://mlsanity.com/paper/2310.06825/) (2023)
- [Training Compute-Optimal Large Language Models](https://mlsanity.com/paper/2203.15556/) (2022)
- [Is Your Code Generated by ChatGPT Really Correct? Rigorous Evaluation of Large Language Models for Code Generation](https://mlsanity.com/paper/2305.01210/) (2023)
- [Instruction-Following Evaluation for Large Language Models](https://mlsanity.com/paper/2311.07911/) (2023)
- [Large Language Monkeys: Scaling Inference Compute with Repeated Sampling](https://mlsanity.com/paper/2407.21787/) (2024)
- [MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models](https://mlsanity.com/paper/2309.12284/) (2023)
- [Explaining Neural Scaling Laws](https://mlsanity.com/paper/2102.06701/) (2021)
- [Scaling Laws for Transfer](https://mlsanity.com/paper/2102.01293/) (2021)
- [The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning](https://mlsanity.com/paper/2312.01552/) (2023)
- [Does Fine-Tuning LLMs on New Knowledge Encourage Hallucinations?](https://mlsanity.com/paper/2405.05904/) (2024)
- [When Scaling Meets LLM Finetuning: The Effect of Data, Model and Finetuning Method](https://mlsanity.com/paper/2402.17193/) (2024)
- [A Careful Examination of Large Language Model Performance on Grade School Arithmetic](https://mlsanity.com/paper/2405.00332/) (2024)
- [The Quantization Model of Neural Scaling](https://mlsanity.com/paper/2303.13506/) (2023)
- [Synthetic Data (Almost) from Scratch: Generalized Instruction Tuning for Language Models](https://mlsanity.com/paper/2402.13064/) (2024)
- [Instruction-tuned Language Models are Better Knowledge Learners](https://mlsanity.com/paper/2402.12847/) (2024)
- [Scaling Laws for Downstream Task Performance of Large Language Models](https://mlsanity.com/paper/2402.04177/) (2024)
- [Why Has Predicting Downstream Capabilities of Frontier AI Models with Scale Remained Elusive?](https://mlsanity.com/paper/2406.04391/) (2024)
- [Do Multi-Hop Question Answering Systems Know How to Answer the Single-Hop Sub-Questions?](https://mlsanity.com/paper/2002.09919/) (2020)
- [Conifer: Improving Complex Constrained Instruction-Following Ability of Large Language Models](https://mlsanity.com/paper/2404.02823/) (2024)
- [Is In-Context Learning Sufficient for Instruction Following in LLMs?](https://mlsanity.com/paper/2405.19874/) (2024)

## Built on by (most cited first, 12 of 12)

- [PivotRL: High Accuracy Agentic Post-Training at Low Compute Cost](https://mlsanity.com/paper/2603.21383/) (2026)
- [Shaping capabilities with token-level data filtering](https://mlsanity.com/paper/2601.21571/) (2026)
- [Data Mixing Optimization for Supervised Fine-Tuning of Large Language Models](https://mlsanity.com/paper/2508.11953/) (2025)
- [What makes Reasoning Models Different? Follow the Reasoning Leader for Efficient Decoding](https://mlsanity.com/paper/2506.06998/) (2025)
- [When Style Breaks Safety: Defending LLMs Against Superficial Style Alignment](https://mlsanity.com/paper/2506.07452/) (2025)
- [ParaVT: Taming the Tool Prior Paradox for Parallel Tool Use in Agentic Video Reinforcement Learning](https://mlsanity.com/paper/2605.20342/) (2026)
- [Balancing the Budget: Understanding Trade-offs Between Supervised and Preference-Based Finetuning](https://mlsanity.com/paper/2502.11284/) (2025)
- [Revisiting Robustness for LLM Safety Alignment via Selective Geometry Control](https://mlsanity.com/paper/2602.07340/) (2026)
- [Operationalising the Superficial Alignment Hypothesis via Task Complexity](https://mlsanity.com/paper/2602.15829/) (2026)
- [LLMs Can Get "Brain Rot": A Pilot Study on Twitter/X](https://mlsanity.com/paper/2510.13928/) (2025)
- [Soft-Prompt Tuning for Fair and Efficient LLM Benchmark Evaluation](https://mlsanity.com/paper/2606.12117/) (2026)
- [YUBI-STAG: Contact and Semantic-Rich Alignment for VLAs via Automated Video-Language Grounding](https://mlsanity.com/paper/2610.09718/) (2026)

## Abstract (arXiv)

The Superficial Alignment Hypothesis posits that almost all of a language model's abilities and knowledge are learned during pre-training, while post-training is about giving a model the right style and format. We re-examine these claims by empirically studying the scaling behavior of post-training with increasing finetuning examples and evaluating them using objective task-specific standardized benchmarks. Through experiments with the Llama-3, Mistral, and Llama-2 model families of multiple sizes, we observe that, similar to the pre-training scaling laws, post-training task performance scales as a power law against the number of finetuning examples. This power law relationship holds across a broad array of capabilities, including mathematical reasoning, coding, instruction following, and multihop-reasoning. In addition, for tasks like math and multihop reasoning, we observe that a handful of examples merely align the model stylistically but do not saturate performance on the benchmarks. Model performance is instead correlated with its reasoning ability and it improves significantly with more examples, illustrating the need for holistic evaluation programs leveraging objective benchmarks in addition to measurement of alignment to human preferences. We also observe that language models are not necessarily limited to using knowledge learned during pre-training. With appropriate post-training, a model's ability to integrate new knowledge greatly improves on downstream tasks like multihop question-answering. Taken together, these results shed new light on the Superficial Alignment Hypothesis, suggesting that it is, at best, an over-simplification.