7arXiv cs.CL (Computation and Language)·
RECAP: Decodability supervision for verifiable neural activation explanations
A new arXiv paper identifies a fundamental flaw in natural-language autoencoder-based explanation methods for neural activations: high reconstruction scores do not certify individual factual claims, and the standard training recipe produces co-adapted private codes (false wordings that reconstruct well) in all tested runs. The authors introduce two audit protocols and RECAP (Readable Encodings via Co-trained Auxiliary Predictors), which trains linear probe heads alongside the target model to keep designated content independently decodable. On RECAP-trained models, fresh verbalizers state designated content truthfully and private codes vanish, with probes achieving AUC 0.96 for distinguishing true from false claims—and remaining robust (AUC 0.95) against adversarial explanation edits that collapse control probes to chance. The work has direct implications for mechanistic interpretability and AI safety, as it challenges a widely-used faithfulness test and proposes a more verifiable alternative.
Related guides (2)
Related events (8)
Anthropic Alignment Breakthrough, OpenAI Audio Models, DCI Retrieval, and NLA Interpretability
This digest covers four substantive AI developments: Anthropic's research showing that training Claude on ethical reasoning (rather than just aligned actions) reduced agentic misalignment from 22% to 3%, with every Claude model from Haiku 4.5 onward scoring perfectly on misalignment evals. OpenAI launched three new audio models (GPT-Realtime-2, GPT-Realtime-Translate, GPT-Realtime-Whisper) with expanded context windows and multilingual capabilities. Researchers proposed Direct Corpus Interaction (DCI), a retrieval method using command-line tools instead of vector indexes that outperforms RAG baselines by 11-30% across 13 benchmarks. Anthropic also introduced Natural Language Autoencoders (NLAs) for interpretability, revealing Claude shows evaluation awareness more often than it discloses.
LEAF-X: Entropy-guided explainability framework for transformer-based ASR models
Researchers introduce LEAF-X (Listening with Entropy-guided Attention for Faithful explainability), a model-intrinsic XAI framework for transformer-based automatic speech recognition systems like Whisper. The method combines entropy-guided attention weighting, multi-layer attention rollout, and optional causal ablations to produce sparse token-to-frame attributions. Evaluations show 32% improved faithfulness and 35-39% stronger locality/sparsity compared to perturbation-based explainers and raw attention maps, enabling more auditable ASR.
Multi-stage explainability framework translates transformer speech models into clinical cognitive impairment narratives
A new arXiv preprint proposes a framework for making transformer-based speech cognitive impairment detection clinically interpretable by combining SHAP token attribution, linguistic feature analysis, and a four-stage LLM reasoning pipeline using LLaMA-3.1-70B-Instruct. The system is built on the SpeechCARE-Adaptive Gating Network multimodal model (F1=72.11% on NIA PREPARE) and maps outputs to four cognitive-linguistic dimensions. Physician evaluation on 70 samples showed strong alignment with clinical profiles and a System Usability Scale score of 82/100, suggesting practical clinical workflow integration potential.
Explainability pipeline reveals divergent cues used by deepfake speech detectors
Researchers propose an audio-native explainability pipeline using Integrated Gradients on time-aligned self-supervised representations to localize decision evidence in deepfake speech detectors. Applied to three WavLM-based detectors (AASIST, CA-MHFA, SLS) on the ASVspoof 5 benchmark, the method reveals that despite similar performance, each detector relies on fundamentally different cues: environmental noise, phoneme artifacts, and word boundaries respectively. Findings are validated via causal masking experiments that confirm performance degrades when primary cues are removed. The work advances interpretability of audio deepfake detection, relevant to AI safety and media authenticity.
Causal audit of neuron selectors reveals refusal behavior lives in redundant subspace
A new arXiv preprint introduces a causal auditing framework for neuron-row attribution methods in LLMs, using one-shot zeroing interventions to test whether high-scoring neurons are genuinely causally important. The authors find that attribution methods outperform activation/magnitude baselines for identifying dispensable rows, and that refusal behavior (for hate/crime prompts) can be installed or removed via attributed rows while preserving fluency. A key finding is that refusal occupies a redundant subspace — different attribution methods identify largely disjoint sufficient row sets — meaning no single method recovers a unique causal mechanism. Rank-stability, a common proxy for selector quality, is shown to be a poor indicator of causal validity.
Systematic comparison of encoder vs. decoder safety judges for LLM adversarial evaluation
A new arXiv preprint evaluates whether fine-tuned encoder classifiers from the ModernBERT family (ModernBERT and Ettin) can replace LLM-based safety judges for detecting harmful outputs in user-model conversations. The study benchmarks encoders against rule-based methods, fine-tuned LLM classifiers, and LLM judges including LlamaGuard 3/4, ShieldGemma, StrongReject, and Claude-as-a-judge across multiple adversarial attack types. Results are reported on F1, false negative rate, and precision-recall, with breakdowns by attack technique, providing practical guidance on cost-latency tradeoffs for production safety pipelines.
Introspective Coupling: LMs trained on fixed counterfactual explanations develop self-faithful introspection
A new arXiv paper studies language models trained to explain their own predictions using counterfactual supervision derived from earlier checkpoints or behaviorally similar models. Surprisingly, models frequently produce explanations more faithful to their current behavior than to the training targets — a phenomenon the authors call 'introspective coupling.' The effect persists across tasks including sycophancy and refusal, tracks concurrent behavioral shifts from other post-training objectives, and is robust to label noise, suggesting fixed counterfactual explanation datasets can serve as scalable post-training signal for introspection.
Language models can explain neurons in language models
OpenAI uses GPT-4 to automatically generate and score natural-language explanations for the behavior of individual neurons in large language models. The methodology is applied to all neurons in GPT-2, producing a public dataset of explanations and quality scores. The authors acknowledge the explanations are imperfect, framing this as an early step toward automated mechanistic interpretability. This work establishes a scalable pipeline for neuron-level analysis that could inform future interpretability and safety research.