Anthropic · Research

Research listings from Anthropic on TechiSeek. 93 public listings.

• Agentic misalignment: How LLMs could be insider threats (Anthropic)
  Agentic misalignment: How LLMs could be insider threats Highlights - We stress-tested 16 leading models from multiple developers in hypothe…
• Paving the way for AI agents in biology (Anthropic)
  In this Anthropic Science post, Laura Luebbert argues that we need to make biological data infrastructure more agent-friendly.
• In-context learning and induction heads (Anthropic)
  We observe a sharp phase change early in training when induction heads form and in-context learning jumps, across models from 13M to 13B pa…
• Assessing Claude Mythos Preview’s cybersecurity capabilities (Anthropic)
  Technical details on how we tested Claude Mythos Preview, a model strikingly capable at computer security tasks, and what we found.
• Automated Alignment Researchers (Anthropic)
  Nine Claude instances worked autonomously on an alignment problem for five days. Results on using language models to scale scalable oversig…
• AI models on realistic cyber ranges (Anthropic)
  Current Claude models can succeed at multistage attacks on networks with dozens of hosts using only standard, open-source tools.
• Building AI for cyber defenders (Anthropic)
  How we've improved Claude's cyber defense skills.
• Coding agents in the social sciences (Anthropic)
  Results from a survey of 1,260 social scientists about AI and coding agent use.
• Towards monosemanticity (Anthropic)
  Using dictionary learning to decompose a transformer layer into thousands of monosemantic features, a path past the polysemantic-neuron pro…
• Persona vectors: Monitoring and controlling character traits in language models (Anthropic)
  A paper from Anthropic describing persona vectors and their applications to monitoring and controlling model behavior.
• Values in the wild: Discovering and analyzing values in real-world language model interactions (Anthropic)
  Values in the wild: Discovering and analyzing values in real-world language model interactions People don’t just ask AIs for the answers to…
• Question decomposition improves reasoning faithfulness (Anthropic)
  Breaking complex questions into subquestions produces more faithful reasoning than chain-of-thought while retaining accuracy gains.
• Anthropic Economic Index: AI's impact on software development (Anthropic)
  Anthropic Economic Index: AI’s impact on software development Jobs that involve computer programming are a small sector of the modern econo…
• Studying large language model generalization with influence functions (Anthropic)
  An EK-FAC approximation makes influence functions orders of magnitude faster on LLMs up to 52B parameters, revealing how models generalize.
• Automated researchers can reliably mitigate alignment failures (Anthropic)
  We had Claude autonomously train models to improve their performance on several public benchmarks that measure 10 categories of alignment f…
• Toy models of superposition (Anthropic)
  Small networks can represent more features than they have dimensions when features are sparse, a phenomenon called superposition.
• How people ask Claude for personal guidance (Anthropic)
  We look at what types of guidance people ask of Claude, and describe how this research shaped the training of our newest models.
• Simple probes can catch sleeper agents (Anthropic)
  Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
• 2028: Two scenarios for global AI leadership (Anthropic)
  Our views on the AI competition between the US and China.
• How Claude is uplifting biomolecular modeling (Anthropic)
  Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
• Project Vend: Can Claude run a small shop? (And why does that matter?) (Anthropic)
  We let Claude run a small shop in the Anthropic office.
• Language models (mostly) know what they know (Anthropic)
  Larger models are well calibrated at judging whether their own answers are true, and can predict whether they know an answer before answeri…
• How AI assistance impacts the formation of coding skills (Anthropic)
  A randomized trial with 52 developers found AI assistance lowered comprehension scores, except when used for explanations rather than deleg…
• A general language assistant as a lab for alignment (Anthropic)
  A general language assistant as a testbed for alignment, where preference modeling outperforms imitation learning and scales better.
• Measuring the persuasiveness of language models (Anthropic)
  Anthropic developed a way to test how persuasive language models (LMs) are, and analyzed how persuasiveness scales across different version…
• A statistical approach to model evaluations (Anthropic)
  Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
• Predictability and surprise in large generative models (Anthropic)
  Aggregate performance of large models is predictable via scaling laws, yet specific capabilities emerge unpredictably.
• Introducing Bloom: Automated behavioral evals (Anthropic)
  Bloom is an open-source agentic framework that auto-generates behavioral evaluations, tested across 16 frontier models.
• A “diff” tool for AI: Finding behavioral differences in new models (Anthropic)
  A new interpretability method compares models with different architectures to surface behavioral differences, tested across four open-sourc…
• Clio: A system for privacy-preserving insights into real-world AI use (Anthropic)
  - This system and research tool is now called Anthropic Insights - Consumer and What do people use AI models for?
• Alignment faking in large language models (Anthropic)
  Most of us have encountered situations where someone appears to share our views or values, but is in fact only pretending to do so—a behavi…
• The persona selection model (Anthropic)
  A theory of why AI models act like humans.
• The assistant axis (Anthropic)
  We investigate the character that most modern language models inhabit when interacting with users.
• Privileged bases in the transformer residual stream (Anthropic)
  Individual coordinates in the residual stream show unexpected privileged basis behavior, provisionally attributed to Adam's per-dimension n…
• Experimenting with AI to defend critical infrastructure (Anthropic)
  Anthropic partnered with Pacific Northwest National Laboratory to explore how AI can help defenders find and fix vulnerabilities before att…
• Sabotage evaluations for frontier models (Anthropic)
  Any industry where there are potential harms needs evaluations. Nuclear power stations have continuous radiation monitoring and regular sit…
• Why do we take LLMs seriously as a potential source of biorisk? (Anthropic)
  This article explains why we believe that evaluating biorisk and safeguarding against it is a critical element of responsible AI developmen…
• Claude’s values across models and languages (Anthropic)
  We analyzed 300,000 real conversations to measure the values Claude expresses across models and languages, compressed into four interpretab…
• Mapping the mind of a large language model (Anthropic)
  We have identified how millions of concepts are represented inside Claude Sonnet, one of our deployed large language models.
• Constitutional AI: Harmlessness from AI feedback (Anthropic)
  Constitutional AI trains a harmless but non-evasive assistant through self-critique and RL from AI feedback, with far fewer human labels.
• Developing nuclear safeguards for AI (Anthropic)
  With the NNSA and DOE national labs, we co-developed a classifier that distinguishes concerning from benign nuclear-related conversations.
• Confidential Inference via Trusted Virtual Machines (Anthropic)
  Every day, millions of users entrust Claude with sensitive information—from proprietary code to confidential business strategies.
• Estimating AI productivity gains (Anthropic)
  What do real conversations with Claude tell us about the effects of AI on labor productivity?