# Faithful Activation Verbalization: Reducing Hallucinations in LLM Representation Interpretation

arXiv: https://arxiv.org/abs/2609.34033 · HTML: https://mlsanity.com/paper/2609.34033/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Preference alignment for recommenders and RAG, one of 10 directions in LLM alignment / RLHF / preference optimisation.
- Builds on 21 of its 30 references that are in the map.
- First posted on arXiv 27 Sep 2026.

## Placed in

- LLM alignment / RLHF / preference optimisation / [Preference alignment for recommenders and RAG](https://mlsanity.com/alignment/preference-alignment-for-recommenders-and-rag/)

## Builds on (most cited first, 21 of 21)

- [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://mlsanity.com/paper/2305.18290/) (2023)
- [Visual Instruction Tuning](https://mlsanity.com/paper/2304.08485/) (2023)
- [BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models](https://mlsanity.com/paper/2301.12597/) (2023)
- [Universal and Transferable Adversarial Attacks on Aligned Language Models](https://mlsanity.com/paper/2307.15043/) (2023)
- [Character-level Convolutional Networks for Text Classification](https://mlsanity.com/paper/1509.01626/) (2015)
- [HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal](https://mlsanity.com/paper/2402.04249/) (2024)
- [JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models](https://mlsanity.com/paper/2404.01318/) (2024)
- [LMSYS-Chat-1M: A Large-Scale Real-World LLM Conversation Dataset](https://mlsanity.com/paper/2309.11998/) (2023)
- [TweetEval: Unified Benchmark and Comparative Evaluation for Tweet Classification](https://mlsanity.com/paper/2010.12421/) (2020)
- [A StrongREJECT for Empty Jailbreaks](https://mlsanity.com/paper/2402.10260/) (2024)
- [Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Models](https://mlsanity.com/paper/2401.06102/) (2024)
- [Text Embeddings Reveal (Almost) As Much As Text](https://mlsanity.com/paper/2310.06816/) (2023)
- [SelfIE: Self-Interpretation of Large Language Model Embeddings](https://mlsanity.com/paper/2403.10949/) (2024)
- [Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers](https://mlsanity.com/paper/2512.15674/) (2025)
- [LatentQA: Teaching LLMs to Decode Activations Into Natural Language](https://mlsanity.com/paper/2412.08686/) (2024)
- [Do Activation Verbalization Methods Convey Privileged Information?](https://mlsanity.com/paper/2509.13316/) (2025)
- [Rep2Text: Decoding Full Text from a Single LLM Token Representation](https://mlsanity.com/paper/2511.06571/) (2025)
- [Universal Activation Verbalizer: A Unified Framework for Cross-Model Activation Explanation](https://mlsanity.com/paper/2605.25903/) (2026)
- [Train the Model, Not the Reader: Decodability Supervision for Verifiable Activation Explanations](https://mlsanity.com/paper/2607.20379/) (2026)
- [Building Better Activation Oracles](https://mlsanity.com/paper/2606.02609/) (2026)
- [Retrieval is Enough: Training-Free Interpretability with a Tool-Using Agent](https://mlsanity.com/paper/2607.16448/) (2026)

## Abstract (arXiv)

Activation verbalization methods such as Activation Oracle and Natural Language Autoencoders decode hidden representations of large language models into human-readable natural language. However, existing methods can produce incomplete or hallucinated descriptions, making their activation verbalizations difficult to trust and use reliably in practice. To this end, we introduce AVPO, a two-stage framework that first reconstructs source text from a hidden activation and then evaluates the resulting text with a separate frozen question-answering model, yielding an explicit and inspectable intermediate readout. We further optimize the inverter with direct preference optimization (DPO), using rewards that capture both semantic recoverability and lexical fidelity. Across six text families, AVPO improves gist- and detail-level information recovery over the strongest baseline by up to 17.1 and 9.3 percentage points, respectively. Crucially, the gains arise from preference optimization rather than fine-tuning on selected reconstructions alone, enabling compact cross-model inverters to surpass donor-matched question-conditioned verbalizers while improving both semantic recoverability and lexical fidelity. Moreover, out-of-distribution case study shows that AVPO better recovers high-level semantics while fabricating fewer details.