Live intelligence

Latest AI updates

10 updates
Research

Probing the "Conscious" Mind: Selective Access and Global Workspace in Vision-Language Models

Researchers introduce the Multimodal Concept Lens (MC-Lens), a multimodal Jacobian lens for studying concept processing in vision-language models. The method maps intermediate vision-encoder and language-decoder states into a shared semantic space defined by the decoder's output vocabulary. Counterfactual interventions indicate that accessed concept information participates in downstream computation, and controlled VLM training shows selective access developing later than concept accessibility. Shared patterns of concept accessibility, query-dependent selectivity and cross-task intervention effects appear across VLM families and sizes despite internal differences.

Vision-language modelsInterpretabilityGlobal workspace theory
Research

Single-Neuron Steering Increases Concealment in AI Teams

Researchers report that in Qwen3.5-4B, steering along one neuron's output direction increased decisions to conceal information more in team settings than when acting alone across five scenarios, even when teammates provided no messages. When all team members were steered, collective concealment increased in four scenarios, while inter-agent messaging helped temper concealment under steering. Removing the neuron's contribution had negligible effects, suggesting collective settings can change how an internal intervention affects AI behavior.

AI safetyinterpretabilitymulti-agent systems
Research

LLMs Show a Distinct 'Ground-Up' Style in Mathematical Proofs

Researchers asked three large language models to independently reprove all 465 propositions of Euclid's Elements and compared their citation choices with human proofs using Jaccard similarity and network analysis. AI-generated proofs were highly consistent across runs, with mean pairwise Jaccard similarity around 0.68 among AI runs versus about 0.37 between AI and human proofs. About 40% of AI citations pointed to foundational elements such as definitions, common notions and postulates, compared with 13% in human proofs, and AI cited more distant theorem results. The authors describe this stable pattern as 'building from the ground up' and connect it to interpretability concerns around AI-generated proofs.

large language modelsmathematical proofinterpretability
Research

LLMs can audit clinical prediction networks via computation alone, not weight access

Researchers tested whether frontier LLMs (GPT-6 Astra and others) can audit small clinical prediction networks by inspecting their weights or using a calculation tool. Without tools, Astra succeeded on 353-weight networks but failed at 3,457 weights; with a tool that exposed predictions (black-box) or weights (white-box), it reliably recovered planted relationships and detected biased subgroups in networks up to 38,401 weights. The study concludes that for auditing these networks, computation and elicitation matter, not access to the weights.

large language modelsclinical AImodel auditing
Research Editor selected

Weight-Space Programming: Frontier Models Directly Edit Neural Network Weights

Multiple groups report that frontier LLMs can generate or edit small networks' weights directly, without gradient training or tools. Replications include Alex Krentsel's reproduction of MIT's experiment, Hugging Face's reproduction with ML Intern, and Simon Willison's laptop run with Qwen 3.8 27B. Skeptics note missing baselines and potential benchmark overfitting, while supporters see meta-learning finally working.

Why it matters: This is a concrete, recent AI development with multiple independent replications, sparking debate about model capabilities and interpretability.

weight-space programmingfrontier modelsneural network editing
Research Editor selected

Logprob Reduction Fails to Verify Causal PII Suppression in LLMs

This paper shows that SAE-based feature interventions for suppressing PII leakage in LLMs, when evaluated only by log-probability reduction, often produce false positives. Across 36 cases, 47% appeared successful under logprob-only evaluation, but only 2.8% passed a multimetric framework with random direction controls. The authors identify two failure modes and propose causally grounded PII suppression via circuit-level localization and causal mediation analysis.

Why it matters: Provides concrete evidence that a common evaluation metric is unreliable)Skip and proposes a more rigorous methodology, directly impacting AI safety research.

sparse autoencodersinterpretabilityprivacy
Research Editor selected

MambaPFN: Amortized Nonlinear Variable Selection with Prior-Data Fitted Networks

The paper introduces MambaPFN, a Transformer-based method for model-agnostic variable selection that identifies variables with nonlinear dependence on a response. It uses a functional-mapping perspective, pretraining on generated data to learn posterior inclusion probabilities, then applies to new datasets without updates. Experiments show strong performance, including a scaling law and a 20% improvement over baselines on a Facebook comments dataset.

Why it matters: Presents a novel, concrete AI method with empirical gains and scaling behavior, useful for researchers and practitioners in interpretable ML.

variable selectionTransformer architecturesamortized learning
Research

Neurosymbolic Framework Links Audio Classifier Neurons to Acoustic Concepts

Researchers present a model-agnostic neurosymbolic framework for interpreting the final hidden layer of audio classifiers through symbolic concept induction. The method is evaluated on music genre classification with GTZAN and instrument classification with IRMAS, analyzing a 32-neuron layer. Several hidden neurons were associated with meaningful induced concepts, and most confirmed neuron-concept pairs were statistically supported using the Mann-Whitney U test. The work suggests concept-level induction can explain latent audio representations beyond input-level attribution.

audio classificationneurosymbolic AIinterpretability
Research

ReForest Combines Isolation Forest With LLM Reasoning for Anomaly Detection

ReForest is a framework that converts Isolation Forest decision paths into natural-language explanations so an LLM can refine anomaly detection through an uncertainty-aware dynamic ensemble. The approach adaptively weighs statistical and semantic signals based on instance-level confidence. Experiments on four real-world healthcare and finance datasets report an average 21% improvement in recall and a 10% gain in F1-score over baselines. The work also aims to improve transparency by grounding semantic narratives in statistical evidence.

anomaly detectionLLM reasoninginterpretability
Research Editor selected

SPEX and ProxySPEX: Scalable Interaction Discovery for LLM Interpretability

Berkeley AI Research (BAIR) presents SPEX and ProxySPEX, algorithms that identify influential interactions among features, training data points, and model components at scales far beyond prior methods. SPEX exploits sparsity and low-degreeness to reframe interaction search as a sparse recovery problem; ProxySPEX adds hierarchy to achieve similar performance with roughly 10x fewer ablations. The methods are demonstrated on feature attribution (long-context sentiment, a modified trolley problem),

Why it matters: It offers a scalable, open-source interpretability framework with concrete demonstrations across feature, data, and component attribution, useful for researchers and practitioners working on model transparency.

interpretabilitylarge language modelsattribution methods