# Training Trajectories Determine Circuit Removability in Annealable Soft-Prior Transformers

arXiv: https://arxiv.org/abs/2609.10287 · HTML: https://mlsanity.com/paper/2609.10287/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Placed in Circuits and mechanistic analysis, one of 13 directions in Mechanistic interpretability.
- Builds on 22 of its 27 references that are in the map.
- First posted on arXiv 9 Sep 2026.

## Placed in

- Mechanistic interpretability / [Circuits and mechanistic analysis](https://mlsanity.com/mechinterp/circuits-and-mechanistic-analysis/)

## Builds on (most cited first, 22 of 22)

- [Attention Is All You Need](https://mlsanity.com/paper/1706.03762/) (2017)
- [Distilling the Knowledge in a Neural Network](https://mlsanity.com/paper/1503.02531/) (2015)
- [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://mlsanity.com/paper/2104.09864/) (2021)
- [The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks](https://mlsanity.com/paper/1803.03635/) (2018)
- [Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small](https://mlsanity.com/paper/2211.00593/) (2022)
- [Self-Attention with Relative Position Representations](https://mlsanity.com/paper/1803.02155/) (2018)
- [Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation](https://mlsanity.com/paper/2108.12409/) (2021)
- [Deep Networks with Stochastic Depth](https://mlsanity.com/paper/1603.09382/) (2016)
- [In-context Learning and Induction Heads](https://mlsanity.com/paper/2209.11895/) (2022)
- [Progress measures for grokking via mechanistic interpretability](https://mlsanity.com/paper/2301.05217/) (2023)
- [Towards Automated Circuit Discovery for Mechanistic Interpretability](https://mlsanity.com/paper/2304.14997/) (2023)
- [What Can Transformers Learn In-Context? A Case Study of Simple Function Classes](https://mlsanity.com/paper/2208.01066/) (2022)
- [What learning algorithm is in-context learning? Investigations with linear models](https://mlsanity.com/paper/2211.15661/) (2022)
- [RepVGG: Making VGG-style ConvNets Great Again](https://mlsanity.com/paper/2101.03697/) (2021)
- [ConViT: Improving Vision Transformers with Soft Convolutional Inductive Biases](https://mlsanity.com/paper/2103.10697/) (2021)
- [Reducing Transformer Depth on Demand with Structured Dropout](https://mlsanity.com/paper/1909.11556/) (2019)
- [Movement Pruning: Adaptive Sparsity by Fine-Tuning](https://mlsanity.com/paper/2005.07683/) (2020)
- [Early Convolutions Help Transformers See Better](https://mlsanity.com/paper/2106.14881/) (2021)
- [The Impact of Positional Encoding on Length Generalization in Transformers](https://mlsanity.com/paper/2305.19466/) (2023)
- [KERPLE: Kernelized Relative Positional Embedding for Length Extrapolation](https://mlsanity.com/paper/2205.09921/) (2022)
- [Contextual Position Encoding: Learning to Count What's Important](https://mlsanity.com/paper/2405.18719/) (2024)
- [DAPE: Data-Adaptive Positional Encoding for Length Extrapolation](https://mlsanity.com/paper/2405.14722/) (2024)

## Built on by (most cited first, 1 of 1)

- [Action Shaping: Policies Absorb What They Can Express](https://mlsanity.com/paper/2609.32752/) (2026)

## Abstract (arXiv)

Soft positional priors can help small Transformers learn retrieval circuits, but it is unclear whether the resulting circuits remain functional once the prior is removed. We test this with an annealable soft-prior Transformer whose attention biases can be learned, faded, or zeroed during training and evaluation. On associative recall, unforced models perform well with the prior active (0.772 ± 0.020) but collapse at zero gate (0.095 ± 0.009). Smooth fade-to-zero training preserves high zero-gate accuracy (0.734 ± 0.028), whereas forced-zero training, hard switching, and post hoc continuation fail to recover the same effect. The pattern also appears on Markov induction. Linear regression ICL provides a boundary case because zero-gate training can learn that task directly. Mechanistic traces show that circuit consolidation occurs after the gate reaches zero, even though the responsible heads vary across seeds. These results suggest that circuit removability in small discrete retrieval tasks depends on the training trajectory, not just the final architecture.