# Fresh in memory: Training-order recency is linearly encoded in language model activations

arXiv: https://arxiv.org/abs/2509.14223 · HTML: https://mlsanity.com/paper/2509.14223/

## Key facts

- Cited by 6 papers in the map (data through 14 Oct 2026).
- Builds on 15 of its 19 references that are in the map.
- First posted on arXiv 17 Sep 2025.

## Builds on (most cited first, 15 of 15)

- [The Llama 3 Herd of Models](https://mlsanity.com/paper/2407.21783/) (2024)
- [LoRA: Low-Rank Adaptation of Large Language Models](https://mlsanity.com/paper/2106.09685/) (2021)
- [Qwen3 Technical Report](https://mlsanity.com/paper/2505.09388/) (2025)
- [Steering Llama 2 via Contrastive Activation Addition](https://mlsanity.com/paper/2312.06681/) (2023)
- [Adafactor: Adaptive Learning Rates with Sublinear Memory Cost](https://mlsanity.com/paper/1804.04235/) (2018)
- [Alignment faking in large language models](https://mlsanity.com/paper/2412.14093/) (2024)
- [Do I Know This Entity? Knowledge Awareness and Hallucinations in Language Models](https://mlsanity.com/paper/2411.14257/) (2024)
- [Probing Representation Forgetting in Supervised and Unsupervised Continual Learning](https://mlsanity.com/paper/2203.13381/) (2022)
- [Manipulating SGD with Data Ordering Attacks](https://mlsanity.com/paper/2104.09667/) (2021)
- [Fortuitous Forgetting in Connectionist Networks](https://mlsanity.com/paper/2202.00155/) (2022)
- [Learned feature representations are biased by complexity, learning order, position, and more](https://mlsanity.com/paper/2405.05847/) (2024)
- [On Linear Representations and Pretraining Data Frequency in Language Models](https://mlsanity.com/paper/2504.12459/) (2025)
- [Extractive Structures Learned in Pretraining Enable Generalization on Finetuned Facts](https://mlsanity.com/paper/2412.04614/) (2024)
- [Implicit meta-learning may lead language models to trust more reliable sources](https://mlsanity.com/paper/2310.15047/) (2023)
- [Reawakening knowledge: Anticipatory recovery from catastrophic interference via structured training](https://mlsanity.com/paper/2403.09613/) (2024)

## Built on by (most cited first, 6 of 6)

- [Why Larger Models Learn More: Effects of Capacity, Interference, and Rare-Task Retention](https://mlsanity.com/paper/2605.29548/) (2026)
- [Optimizer Memory Makes Shuffle Order a First-Order Source of Fine-Tuning Noise](https://mlsanity.com/paper/2606.29554/) (2026)
- [Eliciting Numerical Predictive Distributions of LLMs Without Autoregression](https://mlsanity.com/paper/2603.02913/) (2026)
- [No Reliable Evidence of Self-Reported Sentience in Small Large Language Models](https://mlsanity.com/paper/2601.15334/) (2026)
- [Commutator Memory: Sparse, Path-Local Reading and Steering in Language Models](https://mlsanity.com/paper/2609.34348/) (2026)
- [Conflicting Supervision Moves Commitment, Not Capability: A 12.29σ arrangement effect that is exactly zero under a convention-agnostic score](https://mlsanity.com/paper/2610.00234/) (2026)

## Abstract (arXiv)

We show that language models' activations linearly encode when information was learned during training. Our setup involves creating a model with a known training order by sequentially fine-tuning Llama-3.2-1B on six disjoint but otherwise similar datasets about named entities. We find that the average activations of test samples corresponding to the six training datasets encode the training order: when projected into a 2D subspace, these centroids are arranged exactly in the order of training and lie on a straight line. Further, we show that linear probes can accurately (~90%) distinguish "early" vs. "late" entities, generalizing to entities unseen during the probes' own training. The model can also be fine-tuned to explicitly report an unseen entity's training stage (~80% accuracy). Interestingly, the training-order encoding does not seem attributable to simple differences in activation magnitudes, losses, or model confidence. Our paper demonstrates that models are capable of differentiating information by its acquisition time, and carries significant implications for how they might manage conflicting data and respond to knowledge modifications.