# Final Checkpoints Are Not Enough: Analyzing Latent Reasoning Faithfulness Along Training Trajectories

arXiv: https://arxiv.org/abs/2607.06648 · HTML: https://mlsanity.com/paper/2607.06648/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Builds on 23 of its 32 references that are in the map.
- First posted on arXiv 7 Jul 2026.

## Builds on (most cited first, 23 of 23)

- [Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small](https://mlsanity.com/paper/2211.00593/) (2022)
- [Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting](https://mlsanity.com/paper/2305.04388/) (2023)
- [Show Your Work: Scratchpads for Intermediate Computation with Language Models](https://mlsanity.com/paper/2112.00114/) (2021)
- [Progress measures for grokking via mechanistic interpretability](https://mlsanity.com/paper/2301.05217/) (2023)
- [Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets](https://mlsanity.com/paper/2201.02177/) (2022)
- [Towards Faithfully Interpretable NLP Systems: How should we define and evaluate faithfulness?](https://mlsanity.com/paper/2004.03685/) (2020)
- [Learning the Difference that Makes a Difference with Counterfactually-Augmented Data](https://mlsanity.com/paper/1909.12434/) (2019)
- [Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach](https://mlsanity.com/paper/2502.05171/) (2025)
- [Think before you speak: Training Language Models With Pause Tokens](https://mlsanity.com/paper/2310.02226/) (2023)
- [From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step](https://mlsanity.com/paper/2405.14838/) (2024)
- [CODI: Compressing Chain-of-Thought into Continuous Space via Self-Distillation](https://mlsanity.com/paper/2502.21074/) (2025)
- [How to use and interpret activation patching](https://mlsanity.com/paper/2404.15255/) (2024)
- [Towards Faithful Model Explanation in NLP: A Survey](https://mlsanity.com/paper/2209.11326/) (2022)
- [Faithfulness Tests for Natural Language Explanations](https://mlsanity.com/paper/2305.18029/) (2023)
- [LLM Circuit Analyses Are Consistent Across Training and Scale](https://mlsanity.com/paper/2407.10827/) (2024)
- [LLMs are Single-threaded Reasoners: Demystifying the Working Mechanism of Soft Thinking](https://mlsanity.com/paper/2508.03440/) (2025)
- [Multiple-Choice Questions are Efficient and Robust LLM Evaluators](https://mlsanity.com/paper/2405.11966/) (2024)
- [Do Latent Tokens Think? A Causal and Adversarial Analysis of Chain-of-Continuous-Thought](https://mlsanity.com/paper/2512.21711/) (2025)
- [Implicit Reasoning in Transformers is Reasoning through Shortcuts](https://mlsanity.com/paper/2503.07604/) (2025)
- [Do Latent-CoT Models Think Step-by-Step? A Mechanistic Study on Sequential Reasoning Tasks](https://mlsanity.com/paper/2602.00449/) (2026)
- [Dynamics Within Latent Chain-of-Thought: An Empirical Study of Causal Structure](https://mlsanity.com/paper/2602.08783/) (2026)
- [How Do Latent Reasoning Methods Perform Under Weak and Strong Supervision?](https://mlsanity.com/paper/2602.22441/) (2026)
- [Are Latent Reasoning Models Easily Interpretable?](https://mlsanity.com/paper/2604.04902/) (2026)

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

- [Behaviour Is an Incomplete Measure of Reasoning Development: Cross-surface pre-arrival accessibility and the limits of developmental inference in a recurrent-depth reasoner](https://mlsanity.com/paper/2608.16085/) (2026)

## Abstract (arXiv)

Latent reasoning performs multi-step inference in continuous hidden states, promising more compact and efficient reasoning. However, these opaque states raise a question of faithfulness: whether the latent reasoning steps drive the final answer. Prior work studies this question at selected checkpoints and reports several unfaithful behaviors. This endpoint view leaves how evidence of faithfulness evolves during training unexamined. We track behavioral and activation-based evidence across training using verified counterfactual edits and interventions on the latent reasoning states. We find that high task accuracy can coexist with low counterfactual responsiveness: as accuracy improves, responsiveness can decline, and different latent reasoning approaches follow distinct trajectories. On ProsQA, output sensitivity to norm-noise replacement declines alongside counterfactual responsiveness, although the result depends on the replacement. Across separately trained binary-choice and open-ended GSM settings, intervention sensitivity follows opposite trajectories. These results show that evaluating only a final checkpoint can obscure both when counterfactual responsiveness changes and what the latent states contribute.