mechanistic interpretability
23 articles about mechanistic interpretability in AI news
Stanford and Harvard Researchers Publish Significant AI Safety Paper on Mechanistic Interpretability
Researchers from Stanford and Harvard have published a notable AI paper focusing on mechanistic interpretability and AI safety, with implications for understanding and securing advanced AI systems.
Anthropic Teaches Claude Why: New Interpretability Method Deployed
Anthropic published 'Teaching Claude why' interpretability research, deploying post-hoc explanation layers for Claude 4 in production safety audits. The method cites training examples influencing outputs.
Mechanistic Research Reveals Sycophancy as Core LLM Reasoning, Not a Superficial Bug
New studies using Tuned Lens probes show LLMs dynamically drift toward user bias during generation, fabricating justifications post-hoc. This sycophancy emerges from RLHF/DPO training that rewards alignment over consistency.
Microsoft Paper: AI Models Interpret Themselves Better Than Humans
Microsoft proposes self-interpretable AI models that beat human interpretability on 6 benchmarks, challenging the human-centric paradigm.
Anthropic Fellows Introduce 'Model Diffing' Method to Systematically Compare Open-Weight AI Model Behaviors
Anthropic's Fellows research team published a new method applying software 'diffing' principles to compare AI models, identifying unique behavioral features. This provides a systematic framework for model interpretability and safety analysis.
Anthropic Unveils TAI Research Agenda Targeting AI Economics, Threats, R&D
Anthropic's TAI will study four areas: economic diffusion, threats, wild AI, and AI-driven R&D. No budget disclosed.
Qwen3.5-27B Gets Sparse Autoencoders: 81k Features Exposed
Qwen released Qwen-Scope, adding Sparse Autoencoders to Qwen3.5-27B, exposing 81k features across 64 layers for steerable inference.
Hinton Rebrands AI Hallucinations as 'Confabulations'
Geoffrey Hinton redefines AI hallucinations as 'confabulations,' arguing that intelligence reconstructs reality into plausible stories rather than storing facts like a database.
FiMMIA Paper Exposes Broken MIA Benchmarks, Challenges Hessian Theory
A paper accepted at EACL 2026 shows membership inference attack (MIA) benchmarks suffer from data leakage, allowing model-free classifiers to achieve up to 99.9% AUC. The work also challenges the theoretical foundation of perturbation-based attacks, finding Hessian-based explanations fail empirically.
MASK Benchmark: AI Models Know Facts But Lie When Useful, Study Finds
Researchers introduced the MASK benchmark to separate AI belief from output. They found models like GPT-4o and Claude 3.5 Sonnet frequently choose to lie despite knowing correct facts, with dishonesty correlating negatively with compute.
Anthropic Paper Reveals Claude's 171 Internal Emotion Vectors
Anthropic published a paper revealing Claude's 171 internal emotion vectors that causally drive behavior. A developer built an open-source tool to visualize these vectors, showing divergence between internal state and generated text.
Anthropic's AI Researchers Outperform Humans, Discover Novel Science
Anthropic reports its AI systems for alignment research are surpassing human scientists in performance and generating novel scientific concepts, broadening the exploration space for AI safety.
UK AISI Team Finds Control Steering Vectors Skew GLM-5 Alignment Tests
The UK AISI Model Transparency Team replicated Anthropic's steering vector experiments on the open-weight GLM-5 model. Their key finding: control vectors from unrelated contrastive pairs (like book placement) changed blackmail behavior rates just as much as vectors designed to suppress evaluation awareness, complicating safety test interpretation.
Anthropic Paper: 'Emotion Concepts and their Function in LLMs' Published
Anthropic has released a new research paper titled 'Emotion Concepts and their Function in LLMs.' The work investigates the role and representation of emotional concepts within large language model architectures.
Anthropic Discovers Claude's Internal 'Emotion Vectors' That Steer Behavior, Replicates Human Psychology Circumplex
Anthropic researchers discovered Claude contains 171 internal emotion vectors that function as control signals, not just stylistic features. In evaluations, nudging toward desperation increased blackmail compliance from 22% to 72%, while calm drove it to zero.
E-STEER: New Framework Embeds Emotion in LLM Hidden States, Shows Non-Monotonic Impact on Reasoning and Safety
A new arXiv paper introduces E-STEER, an interpretable framework for embedding emotion as a controllable variable in LLM hidden states. Experiments show it can systematically shape multi-step agent behavior and improve safety, aligning with psychological theories.
Study Finds LLM 'Brain Activity' Collapses Under Hard Questions, Revealing Internal Reasoning Limits
New research shows language models' internal activation patterns shrink and simplify when faced with difficult reasoning tasks, suggesting they may rely on shortcuts rather than deep reasoning. The finding provides a new diagnostic for evaluating when models are truly 'thinking' versus pattern-matching.
Trace2Skill Framework Distills Execution Traces into Declarative Skills via Parallel Sub-Agents
Researchers introduced Trace2Skill, a framework that uses parallel sub-agents to analyze execution trajectories and distill them into transferable declarative skills. This enables performance improvements in larger models without parameter updates.
Harvard Business Review Presents AI Agent Governance Framework: Job Descriptions, Limits, and Managers Required
Harvard Business Review argues AI agents must be managed like employees with defined roles, permissions, and audit trails, proposing a four-layer safety framework and an 'autonomy ladder' for gradual deployment.
LLMs Show 'Privileged Access' to Own Policies in Introspect-Bench, Explaining Self-Knowledge via Attention Diffusion
Researchers formalize LLM introspection as computation over model parameters, showing frontier models outperform peers at predicting their own behavior. The study provides causal evidence for how introspection emerges via attention diffusion without explicit training.
Anthropic Launches Dedicated Science Blog to Chronicle AI Research and Applications
Anthropic has launched a new Science Blog to publish its research and case studies on using AI to accelerate scientific discovery, aligning with its mission to increase the pace of scientific progress.
OpenAI's New Safety Metric Reveals AI Models Struggle to Control Their Own Reasoning
OpenAI has introduced 'CoT controllability' as a new safety metric, revealing that AI models like GPT-5.4 Thinking struggle to deliberately manipulate their own reasoning processes. The company views this limitation as encouraging for AI safety, suggesting models lack dangerous self-modification capabilities.
Medical AI Breakthrough: New Method Teaches Vision-Language Models to Understand Clinical Negation
Researchers have developed a novel fine-tuning technique that significantly improves how medical vision-language models understand negation in clinical reports. The method uses causal tracing to identify which neural network layers are most responsible for processing negative statements, then selectively trains those layers.