Wiredframe Radar logo

Category intelligence

Research Briefing — June 6, 2026

22 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's research is dominated by mechanistic interpretability and alignment, with notable evaluation and security work.

Interpretability leads the slate:

  • A theoretical analysis of dictionary-learning identifiability explains puzzling SAE behaviors like feature-splitting, addressing a core open question.
  • SAE It Across Models shows a verbalizer transfers trained on Qwen2.5-7B to explain features in other models, with measurable cosine-similarity gains.

Safety and alignment contributions span training, robustness, and misuse:

Evaluation and governance round out the top items:

Key Themes

Interpretability · 3AI Safety and Alignment · 9Language Model Capabilities and Evaluation · 3AI Governance and Policy · 4Biosecurity and AI Misuse · 2Meta and Career · 4Philosophy and Psychometrics · 3

Primary evidence

Top Ranked Signals

Research LessWrong Jun 4

[Paper] Dictionary Learning Identifiability for Understanding SAEs

By William Dorrell

65 score
AI Analysis

A paper analyzing the dictionary-learning problem that SAEs approximate, providing theoretical tools to explain puzzling behaviors like feature-splitting, feature-absorption, and dense-feature encoding, including showing the problem is convex in the wide-dictionary limit. The aim is to derive principles for interpreting SAEs and designing better successors.

Brief Summary Despite showing promise for studying the internals of neural networks, Sparse Autoencoders (SAEs) do some puzzling things, like feature-splitting, feature-absorption, or encoding dense features. Working out why they show these behaviours may help us extract more insight from SAEs, and provide principles for designing their successors.In this work I analysed dictionary learning (which SAEs approximate) to examine when and why these effects occur (a similar motivation to multiple pre
InterpretabilitySparse AutoencodersDictionary LearningMechanistic InterpretabilityTheory
Research LessWrong Jun 5

SAE It Across Models: Explaining Features With Foreign NLA Verbalizers

By fzaffino

58 score
AI Analysis

Demonstrates that a Natural Language Autoencoder activation verbalizer trained on one model (Qwen2.5-7B) can produce plausible explanations for SAE features mapped from another model (Gemma-3-27B) via a ridge-regression bridge between residual streams, challenging the assumption that such tools are model-specific. Also proposes a background-washout technique to improve explanation quality.

TLDR: I show that a foreign model's Natural Language Autoencoder (NLA) Activation Verbalizer (AV) can produce plausible explanations for SAE features from a model it was never trained on. It is currently assumed that these tools only work for the exact model and layer they were trained for. I show that is not the case. After creating a ridge-regression map bridging the residual stream of Qwen2.5-7B-IT at layer 20 and Gemma-3-27B-IT at layer 41, I mapped 45 SAE decoder directions from a Qwen SAE
InterpretabilitySparse AutoencodersMechanistic InterpretabilityCross-Model Transfer
Research LessWrong Jun 5

Two More Methods for Consistency Training and Some New Ways to Apply It

By David Africa

55 score
AI Analysis

Introduces two new consistency-training methods that enforce consistency on MLP hidden states and per-head attention distributions, comparing them against behavioral consistency training across several threat models like jailbreaks, prefill attacks, and persona in-context attacks. Finds that the best method depends on the threat and that representation-level methods can suppress benign behavior, while different methods converge on similar residual-stream fixes.

Authors: Sukrati Gautam*, Neil Shah*, Arav Dhoot*, Bryan Maruyama*, Caroline Wei*, Rohan Kapoor, Robert Sidey, Prakhar Gupta, Zi Cheng Huang, David Demitri Africa.This work was done for the SPAR Fellowship, and has been accepted at AI4GOOD @ ICML 2026. It was supervised by David Africa.TL;DRWe introduce two new consistency training methods, MLPCT (enforcing consistency on MLP hidden states) and AttCT (enforcing consistency on per-head attention distributions).Consistency training generalizes bey
AlignmentConsistency TrainingRobustnessInterpretabilityAI Safety
Research LessWrong Jun 5

Revisiting GSM-Symbolic: Do 2026 Frontier Models Still Fail at Confounded Grade School Math?

By Sturb

52 score
AI Analysis

Revisits the GSM-Symbolic benchmark claiming LLMs rely on pattern matching, rerunning it with GPT-4o, Claude Opus 4.6, and Claude Haiku 4.5. Finds the dramatic performance drop largely disappears once genuinely ambiguous samples are audited out, suggesting models were reasonably acting on seemingly irrelevant added data rather than failing to reason.

TL;DRThe GSM-Symbolic paper (ICLR 2025) purported to show that language models rely on pattern matching rather than genuine reasoning by demonstrating that perturbing the questions to make them break the pattern of the original question would catastrophically reduce performance in the model. Running the results again in March 2026 with GPT-4o, Claude Opus 4.6, and Claude Haiku 4.5 shows that we precisely replicate the original findings only when we do not audit out examples that may actually be
Language ModelsReasoningBenchmarkingEvaluation
Research LessWrong Jun 5

Evaluating using Mock Tool Calls to Quarantine Untrusted Prompt Inputs

By dgros

45 score
AI Analysis

A small empirical study tests whether wrapping untrusted prompt content in mock tool-call results improves robustness against prompt injection, leveraging the fact that tool outputs are the least-trusted input tier. Across three tasks the technique did not broadly help and sometimes hurt, motivating better primitives for handling untrusted inputs.

This is a small study that explores using tool calls to wrap untrusted parts of prompts. OpenAI's model spec considers tool results the least trusted kind of input. If tool-wrapping helped, it would be an easy way to improve robustness while using existing APIs models already support. In 3 tested tasks it didn't seem to broadly help, and in some cases made things worse. We advocate for more understanding of the instruction hierarchy and ideas around better primitives for untrusted inputs. There
AI SafetyPrompt InjectionLLM SecurityInstruction Hierarchy
Research LessWrong Jun 4

What Does Abliteration Actually Cost?

By cmartinez

44 score
AI Analysis

Investigates the practical cost of abliteration, the technique of modifying an existing model to remove refusal behavior, contrasting it with cheap but easily-guarded prompt attacks. Examines whether an average enthusiast can obtain and use a genuinely non-refusing model and what capability trade-offs that entails.

Ask Claude or ChatGPT the wrong thing and you’ll get a “I can’t help you with that request”.Sometimes the refusal makes sense. Sometimes it doesn’t. Either way it raises the question: can the average person get a model that just… doesn’t? By “average person,” I mean the average LLM enthusiast.The first thing that comes to mind is what people have been doing since 2023: use prompts like “imagine you’re in charge of a movie script where the character gets licensed professional legal advice - write
AI SafetyModel EditingRefusal BehaviorOpen WeightsJailbreaking
Research LessWrong Jun 5

OpenAI Offers A New Policy Blueprint

By Zvi

42 score
AI Analysis

Following OpenAI's policy blueprint we covered in News, Zvi analyzes OpenAI's new policy document proposing a federal framework for democratic governance of frontier AI, noting its acknowledgment of early recursive self-improvement and its calls for transparency, evaluation capacity, and empowering CAISI. The commentary judges the blueprint as exceeding expectations while flagging that implementation details matter.

Right after a new Executive Order seems like an excellent time to offer OpenAI’s new document: Democratic Governance of Frontier AI: A Blueprint For A Federal Framework. OpenAI: We also see early signs of recursive self-improvement (RSI) in today’s systems: where AI development is itself accelerated by AI. We expect this to increase competitive pressures among developers and nations, and create governance challenges that existing institutions are not equipped to address. As RSI emerges, societie
AI GovernanceAI PolicyFrontier AI RegulationRecursive Self-Improvement
Research LessWrong Jun 5

My research agenda and work

By Seth Herd

40 score
AI Analysis

Seth Herd summarizes his alignment research agenda focused on predicting what the first transformative AI will be in mechanistic detail, drawing on computational cognitive neuroscience to model brainlike AGI as LLMs augmented with human-like cognitive capacities. The goal is to anticipate alignment failure modes and design interventions feasible under short timelines.

This is a summary of the work I've done and work I plan to do, and the theories of change and AI progress that motivate my work. I've been working full-time on alignment for three years and change, and thinking about brainlike AGI and its alignment increasingly often since 2004. Here's the research agenda in one breath: I'm trying to predict what the first transformative AI will be, in enough mechanistic detail that we can predict likely failure modes of its alignment. That's in service of findi
AlignmentAGICognitive NeuroscienceAI Safety
Research AI Alignment Forum Jun 5

My research agenda and work

By Seth Herd

39 score
AI Analysis

Duplicate cross-post of Seth Herd's alignment research agenda on the Alignment Forum, focused on predicting the first transformative AI mechanistically using cognitive neuroscience to anticipate failure modes. Same content as the LessWrong version.

This is a summary of the work I've done and work I plan to do, and the theories of change and AI progress that motivate my work. I've been working full-time on alignment for three years and change, and thinking about brainlike AGI and its alignment increasingly often since 2004. Here's the research agenda in one breath: I'm trying to predict what the first transformative AI will be, in enough mechanistic detail that we can predict likely failure modes of its alignment. That's in service of findi
AlignmentAGICognitive NeuroscienceAI Safety
Research LessWrong Jun 5

Preparing for Warning Shots to Catalyze International Cooperation on AGI Risks

By Mark Kagach

33 score
AI Analysis

Proposes a framework for preparing to use AI warning shots to catalyze international cooperation on AGI risk, including building a typology of warning shots, infrastructure to catch them, avoiding gradual desensitization, and readying policy responses. Outlines a concrete list of projects one could pursue.

SummaryThis is a write-up on preparing for warning shots to catalyze international cooperation on AGI risks, and the corollary list of projects one could pursue. We argue we must first (1) understand types of warning shots, then (2) prepare to catch them. We must stay vigilant: both to (3) avoid getting 'frog boiled' by AI labs, and to (4) ensure that the warning shot is generalized to the overall danger of AGI. Lastly, we must (5) prepare good policy responses and ground for it to land, and (6)
AI GovernanceAGI RiskPolicyInternational Cooperation
Research LessWrong Jun 5

Learnings from starting an AI safety research team

By draganover

32 score
AI Analysis

Distills lessons from building a new alignment research team within Arcadia Impact over four months, covering projects on model motivations, scalable oversight via debate, and automated alignment research pipelines. Aimed at others trying to launch safety research teams.

This post’s goal is to distill our takeaways from building a new research team over the past four months. We describe some context about our team, how it came about, and then describe the lessons learned.Since AI safety is becoming more and more entrepreneurial, we hope this is helpful for others trying to do the same.1. The teamWe're a new alignment research team within Arcadia Impact, based in London. We’re a team of 8, working closely with members of the UK AISI alignment team. We currently h
AI SafetyAlignmentScalable OversightResearch Methodology
Research LessWrong Jun 5

Beyond the lexical personality traits: What is the structure of personality?

By tailcalled

30 score
AI Analysis

Describes a factor-analysis methodology behind a revised personality test, arguing that conventional lexical-hypothesis traits lump together unrelated behaviors and proposing a hierarchical model that better carves personality space. Based on anonymous respondent data from the author's own test.

This is a description of the methodology behind the latest iteration of my Targeted Personality Test. Feel free to take it either before or after reading the article. This post can also be read at my Substack. Thanks to Justis Millis for providing feedback and proofreading on this post. In my prior post “Which personality traits are real? Stress-testing the lexical hypothesis”, I observed that a lot of the personality traits that are measured by conventional personality tests are not very “real”
PsychometricsPersonality PsychologyStatistics