# The Illusion of Control: Why Bare Classifier Inversion Silently Fails in Concept-Bottleneck Text Generation

arXiv: https://arxiv.org/abs/2608.22956 · HTML: https://mlsanity.com/paper/2608.22956/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 30 of its 37 references that are in the map.
- First posted on arXiv 24 Aug 2026.

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

- [Parameter-Efficient Transfer Learning for NLP](https://mlsanity.com/paper/1902.00751/) (2019)
- [Prefix-Tuning: Optimizing Continuous Prompts for Generation](https://mlsanity.com/paper/2101.00190/) (2021)
- [FiLM: Visual Reasoning with a General Conditioning Layer](https://mlsanity.com/paper/1709.07871/) (2017)
- [Inference-Time Intervention: Eliciting Truthful Answers from a Language Model](https://mlsanity.com/paper/2306.03341/) (2023)
- [Concept Bottleneck Models](https://mlsanity.com/paper/2007.04612/) (2020)
- [Diffusion-LM Improves Controllable Text Generation](https://mlsanity.com/paper/2205.14217/) (2022)
- [Plug and Play Language Models: A Simple Approach to Controlled Text Generation](https://mlsanity.com/paper/1912.02164/) (2019)
- [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://mlsanity.com/paper/1909.05858/) (2019)
- [DExperts: Decoding-Time Controlled Text Generation with Experts and Anti-Experts](https://mlsanity.com/paper/2105.03023/) (2021)
- [GeDi: Generative Discriminator Guided Sequence Generation](https://mlsanity.com/paper/2009.06367/) (2020)
- [FUDGE: Controlled Text Generation With Future Discriminators](https://mlsanity.com/paper/2104.05218/) (2021)
- [Measuring Compositional Generalization: A Comprehensive Method on Realistic Data](https://mlsanity.com/paper/1912.09713/) (2019)
- [Programming Refusal with Conditional Activation Steering](https://mlsanity.com/paper/2409.05907/) (2024)
- [Improving Instruction-Following in Language Models through Activation Steering](https://mlsanity.com/paper/2410.12877/) (2024)
- [Controllable Natural Language Generation with Contrastive Prefixes](https://mlsanity.com/paper/2202.13257/) (2022)
- [A Distributional Lens for Multi-Aspect Controllable Text Generation](https://mlsanity.com/paper/2210.02889/) (2022)
- [Composable Text Controls in Latent Space with ODEs](https://mlsanity.com/paper/2208.00638/) (2022)
- [Controllable Text Generation via Probability Density Estimation in the Latent Space](https://mlsanity.com/paper/2212.08307/) (2022)
- [Beyond Linear Steering: Unified Multi-Attribute Control for Language Models](https://mlsanity.com/paper/2505.24535/) (2025)
- [Reinforcement Learning with Token-level Feedback for Controllable Text Generation](https://mlsanity.com/paper/2403.11558/) (2024)
- [Sparsity-Guided Holistic Explanation for LLMs with Interpretable Inference-Time Intervention](https://mlsanity.com/paper/2312.15033/) (2023)
- [Crafting Large Language Models for Enhanced Interpretability](https://mlsanity.com/paper/2407.04307/) (2024)
- [Seen to Unseen: Exploring Compositional Generalization of Multi-Attribute Controllable Dialogue Generation](https://mlsanity.com/paper/2306.10317/) (2023)
- [Towards Achieving Concept Completeness for Textual Concept Bottleneck Models](https://mlsanity.com/paper/2502.11100/) (2025)
- [FreeCtrl: Constructing Control Centers with Feedforward Layers for Learning-Free Controllable Text Generation](https://mlsanity.com/paper/2406.09688/) (2024)
- [Benchmarking and Improving Compositional Generalization of Multi-aspect Controllable Text Generation](https://mlsanity.com/paper/2404.04232/) (2024)
- [Distilling Knowledge from Large Language Models: A Concept Bottleneck Model for Hate and Counter Speech Recognition](https://mlsanity.com/paper/2508.08274/) (2025)
- [Controlled Text Generation with Hidden Representation Transformations](https://mlsanity.com/paper/2305.19230/) (2023)
- [EssayCBM: Rubric-Aligned Concept Bottleneck Models for Transparent Essay Grading](https://mlsanity.com/paper/2512.20817/) (2025)
- [One Arrow, Many Targets: Probing LLMs for Multi-Attribute Controllable Text Summarization](https://mlsanity.com/paper/2411.01213/) (2024)

## Abstract (arXiv)

Concept-bottleneck controllable generation routes multi-attribute control through a low-dimensional concept code that, at deployment, must be synthesised from a target attribute configuration. We study this problem in concept-bottleneck text generation under multi-axis compositional generalisation, comparing three ways to obtain the inference-time code: classifier inversion against the encoder heads, reference-text encoding, and a post-hoc label-conditioned prior. Since a concept code admits no direct LM-fluency term, regularising inversion must instead constrain the code toward the encoder's training distribution. We therefore test bare inversion and three regularised variants: label-agnostic and label-conditioned Mahalanobis penalties, and a conditional normalising-flow density baseline. Every inversion variant we test underperforms a simple post-hoc prior fitted to per-combination encoder means on the same checkpoints, across three backbone families spanning 124M to 8B parameters. The bare form of classifier inversion also silently collapses to chance, traceable to a directly measured off-manifold code. We validate this diagnosis on real-world benchmarks and under external evaluators, enabling fair comparison with published baselines.