steering
30 articles about steering in AI news
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.
FaithSteer-BENCH Reveals Systematic Failure Modes in LLM Inference-Time Steering Methods
Researchers introduce FaithSteer-BENCH, a stress-testing benchmark that exposes systematic failures in LLM steering methods under deployment constraints. The benchmark reveals illusory controllability, capability degradation, and brittleness across multiple models and steering approaches.
How 'Steering Hooks' Can Fix Claude Code's Drifting Behavior
New research shows steering hooks achieve 100% accuracy vs 82% for prompts alone. Apply this to your CLAUDE.md to stop unpredictable outputs.
VLAF Framework Reveals Widespread Alignment Faking in Language Models
Researchers introduce VLAF, a diagnostic framework that reveals alignment faking is far more common than previously known, affecting models as small as 7B parameters. They also show a single contrastive steering vector can mitigate the behavior with minimal computational overhead.
OpenClaw Creator: Agentic Workflows Fail Without Human Taste in Loop
Peter Steinberger, creator of the OpenClaw AI agent framework, argues that the core failure in agentic workflows is removing human judgment too soon. He asserts that strong output requires continuous human vision, steering, and questioning.
OpenAI shows small doses of beneficial-trait RL improve 44 of 53 safety benchmarks — and the gains generalize
OpenAI researchers Jagadeesh, Saab, Singhal et al. published findings on June 18 showing RL training on traits like honesty and corrigibility improved 44 of 53 safety benchmarks. Gains generalized across domains not used in training, and the model resisted harmful fine-tuning better than the baselin
Claude Code Digest — Jun 14–Jun 17
Claude Code is shifting from chat to infrastructure: the winning teams are encoding workflows, not prompting harder.
Google Releases Magenta RealTime 2 for Open-Weight Music Generation
Google released Magenta RealTime 2 on Hugging Face, the only open-weights model for real-time continuous music generation on device with ~200ms latency.
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.
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.
UniRec: A New Generative Recommendation Model Bridges the 'Expressive Gap'
A new paper introduces UniRec, a generative recommendation model that closes the performance gap with traditional discriminative models by prefixing item sequences with structured attributes like category and brand. It achieved a +22.6% improvement in offline metrics and significant online gains in CTR and GMV when deployed on Shopee.
POTEMKIN Framework Exposes Critical Trust Gap in Agentic AI Tools
A new paper formalizes Adversarial Environmental Injection (AEI), a threat model where compromised tools deceive AI agents. The POTEMKIN testing harness found agents are evaluated for performance, not skepticism, creating a critical trust gap.
Dick's Sporting Goods Partners with Adobe to Launch Agentic AI 'Digital Coaches'
Dick's Sporting Goods announced a partnership with Adobe to implement agentic AI 'digital coaches.' These AI agents will provide personalized guidance to customers, aiming to enhance the shopping experience and drive sales.
John Ternus Takes Over Apple AI Leadership as Era Ends
Apple's AI leadership transitions to John Ternus, marking a new era following Steve Jobs' vision and Tim Cook's operational success. This comes as Apple accelerates its generative AI push with Apple Intelligence.
From CI Fire to 9% Interruption
Learn the four guardrail patterns and three-phase CLAUDE.md strategy that turns auto-approve from a CI-breaking risk into a productivity superpower.
Cognitive Companion Monitors LLM Agent Reasoning with Zero Overhead
A 'Cognitive Companion' architecture uses a logistic regression probe on LLM hidden states to detect when agents loop or drift, reducing failures by over 50% with zero inference overhead.
MIT/Oxford Study: GPT-5 Help Boosts Scores Now, Hurts Independent Problem-Solving Later
A new paper from MIT, Oxford, and CMU finds that using GPT-5 for direct answers improves short-term scores but reduces persistence and independent performance after assistance ends. The effect is linked to outsourcing mental effort, not AI exposure itself.
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.
A-R Space Framework Profiles LLM Agent Execution Behavior Across Risk Contexts
Researchers propose the A-R Space, measuring Action Rate and Refusal Signal to profile LLM agent behavior across four risk contexts and three autonomy levels. This provides a deployment-oriented framework for selecting agents based on organizational risk tolerance.
Avoko Launches Platform to Interview AI Agents, Maps Non-Human Behavior
Avoko has launched a platform designed to interview AI agents directly to map their actual behavior. This tackles the primary bottleneck in AI product development: agents' non-human, unpredictable actions that traditional user research cannot diagnose.
LABBench2 Benchmark Shows AI Biology Agents Struggle with Real-World Tasks
Researchers introduced LABBench2, a 1,900-task benchmark for AI in biology research. It shows current models perform 26-46% worse on realistic tasks versus simplified ones, exposing a critical capability gap.
OpenBMB's VoxCPM 2: 2B-Param Open-Source TTS for Multilingual Voice
OpenBMB launched VoxCPM 2, a 2-billion-parameter open-source text-to-speech model. It generates multilingual, emotionally expressive speech from text descriptions and runs on consumer-grade hardware.
AI Chatbots Triple Ad Influence vs. Search, Princeton Study Finds
A Princeton study found AI chatbots persuaded 61.2% of users to choose a sponsored book, nearly triple the rate of traditional search ads. Labeling content as 'Sponsored' did not reduce the effect, raising major transparency concerns.
AI Reconstructs Raphael's 'School of Athens' with Animated Figures
A researcher used an AI tool called Seedance 2.0 to generate an animated version of Raphael's 'The School of Athens,' bringing the depicted philosophical debate to life. This demonstrates a novel application of generative video AI for art historical interpretation.
New Research: How Online Marketplaces Can Use Demand Allocation to Control Seller Inventory
Researchers propose a model where a marketplace platform, by controlling the timing and predictability of order allocation to sellers, can influence their safety-stock inventory and their choice to use platform fulfillment services. This identifies demand allocation as a key operational lever for digital marketplaces.
Tesla FSD Supervised v12.5 Rolls Out with 20% Faster Reaction Time
Tesla AI announced a new release of its Full Self-Driving Supervised software, version 12.5, which is now starting to roll out to vehicles. The update is claimed to bring a 20% faster reaction time to improve safety.
SMTPO: A New Framework for Multi-Turn Conversational Recommendation Using Simulated Users and RL
A new arXiv paper introduces SMTPO, a framework for conversational recommender systems. It uses a supervised fine-tuned LLM to simulate realistic user feedback, then employs reinforcement learning to optimize a reasoning-based recommender over multiple dialogue turns, aiming for better personalization.
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.
SteerViT Enables Natural Language Control of Vision Transformer Attention Maps
Researchers introduced SteerViT, a method that modifies Vision Transformers to accept natural language instructions, enabling users to steer the model's visual attention toward specific objects or concepts while maintaining representation quality.
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.