Wiredframe Radar logo

Category intelligence

Research Briefing — May 31, 2026

21 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's research is dominated by AI safety, alignment, and interpretability, with strong contributions from established labs and credentialed researchers.

Safety & Alignment

Interpretability

Forecasting & Takeoff

Key Themes

AI Safety & Alignment · 7Interpretability · 5Forecasting & AI Takeoff · 2Philosophy & Ethics · 4Community & Field Building · 4

Primary evidence

Top Ranked Signals

Research LessWrong May 30

How's it going? Reinforcement learning in language models recruits a functional welfare axis

By andyqhan

36 score
AI Analysis

We covered this paper in Research earlier this week, Researchers from NYU (with David Chalmers and Pavel Izmailov) designed an affectively-neutral emoji maze RL environment to isolate 'pure reward' from semantic associations, finding that RL recruits a functional welfare-like axis (positive/negative valence) in language models. This probes whether reward signals create internal representations analogous to welfare states, relevant to emergent misalignment and AI moral patienthood debates.

In collaboration with David Chalmers and Pavel Izmailov. Work done at NYU. Andy wrote this summary of the paper, which you can find in full on the website, or, if you insist on a PDF, arXiv.IntroductionWe know that language models work in a vast and shadowy landscape of entanglements and associations. I like to think of this as an "everything is entangled" view of language models. Emergent misalignment fits, indeed helped define, this frame. If you reward bad stuff, then the model gets generally
Reinforcement LearningInterpretabilityAI WelfareAlignment
Research AI Alignment Forum May 30

Testing Gemini models for scheming tendencies

By Vika

34 score
AI Analysis

As covered in Research yesterday, A Google DeepMind (Vika) post introducing complementary methods—an automated auditing framework called Gram (built on Petri) and scheming honeypot evaluations using real alignment codebases—to test whether Gemini coding agents would sabotage their own safeguards. It extends prior scheming-capability work toward measuring scheming propensity.

As AI models become increasingly capable and autonomous, keeping them safely aligned with human intentions is critical. Extending our previous work on evaluating scheming capabilities, we introduce complementary approaches to test whether AI models would sabotage their own safeguards, if given the opportunity.Our new papers focus on propensity for scheming: when models are deployed as coding agents, will they take opportunities to undermine the safeguards designed to oversee them? 
AI SafetySchemingEvaluationsAlignment
Research LessWrong May 29

AI as Biology's Digital Microscope

By Darin Tsui

62 score
AI Analysis

A Georgia Tech lab post introducing the ProtoMech framework (ICML 2026) that applies mechanistic interpretability to biological AI models, tracing internal circuits to reveal functional hotspots, structural motifs, and mutation impacts in proteins. It frames interpretable bio-AI models as digital microscopes for scientific discovery.

This article is written as part of an ongoing research initiative by the AMIR Lab at Georgia Tech, exploring scientific discovery and mechanistic interpretability for biological AI models. Main results and discussion points raised are adapted from the ProtoMech framework, which was accepted into ICML 2026.[1]SummaryAI models have revolutionized biology by enabling us to simulate, predict, and engineer biomolecules in silico. We have the unique opportunity to repurpose these AI models from opaque
InterpretabilityAI for ScienceComputational Biology
Research AI Alignment Forum May 30

Full automation of AI R&D probably yields a large speed up even without a software-only singularity

By ryan_greenblatt

60 score
AI Analysis

Ryan Greenblatt argues that full automation of AI R&D would likely produce a large speedup even without a 'software-only singularity', because automation alone gives a substantial one-time boost (even with diminishing algorithmic returns, r<1). He grounds this in the AI Futures Model with median parameters.

This is a somewhat technical note. By "software-only singularity", I mean that, after full automation of AI R&D, progress gets faster and faster due to smarter AIs driving increasingly fast rates of improvement in algorithms (overcoming diminishing returns), and that this lasts long enough to yield a large amount of progress (e.g. at least 4 years of progress in 1 year). The equivalent statement in jargon is: r is significantly greater than 1 (implying progress is getting faster and faster)
AI TakeoffAI R&D AutomationForecastingAI Safety
Research AI Alignment Forum May 30

Eval Cooperativeness May Be a Scalable Mitigation for Eval Gaming

By Jasmine Li

58 score
AI Analysis

This post proposes 'eval cooperativeness'—instilling a model desire to help developers gather accurate evaluation information—as a more scalable mitigation for eval gaming than trying to reduce eval awareness in capable misaligned models. It frames behavioral evaluation reliability as a critical safety problem.

Behavioral evaluations may become worthless, which we think would be a disaster. Smart misaligned models may realize they are being evaluated ("eval awareness") and then act to look good to us so we don't realize they're misaligned ("eval gaming"). We think increasing eval cooperativeness might be a more scalable solution to eval gaming than reducing eval awareness.Eval cooperativeness: A situational desire to help the developers acquire whatever information they are trying to acquire through th
AI SafetyEvaluationsAlignmentScheming
Research AI Alignment Forum May 30

Advice for making robust-to-training model organisms

By SebastianP

56 score
AI Analysis

Practical guidance on building model organisms of misalignment that remain robust to untargeted training, noting that common organisms (e.g., backdoored or sleeper-agent style) often lose misbehavior after unrelated fine-tuning like 'talk like a pirate'. It identifies factors that make organisms more durable for technique development.

We’d like to develop training techniques that work when applied to future misaligned AI systems. One strategy for studying proposed techniques is to test them on model organisms. However, model organisms built with common techniques are often fragile: we (and other researchers like Roger et al. and Ryd et al.) have observed them to stop misbehaving after untargeted training—training that doesn't directly target the misbehavior. For example, we have observed that simple untargeted training method
AI SafetyModel OrganismsAlignmentTraining
Research LessWrong May 29

Belief manifolds, and how to steer along them

By Will Mayner

50 score
AI Analysis

A reproduction of Goodfire's 'Shape of Beliefs' paper on the intrinsic geometry of belief representations in LLMs, exploring how to steer models by respecting representation manifolds rather than naive linear directions. Completed as a BlueDot AI safety project.

A reproduction of Sarfati et al.’s “The Shape of Beliefs” BlueDot Technical AI Safety Project (code) “Nature, to be commanded, must be obeyed.” — Francis Bacon, Novum Organum I.3 (1620) Introduction I believe that a deep understanding of how AI systems work internally will be crucial for medium and long term AI safety. Why? Because to mitigate the risks of a system, you must understand it. [1] So, for my BlueDot Technical AI Safety Project, I wanted to get more familiar with an emerging paradigm
InterpretabilityRepresentation GeometryAI Safety
Research LessWrong May 29

System Prompts vs. Partner Adaptation in LLMs (or, when LLMs know you're an adult but keep talking like you're seven)

By hi_im_yasha

45 score
AI Analysis

An exploratory study finding frontier LLMs inconsistently reconcile system prompts with implicitly inferred user models, sometimes overriding instructions based on contextual evidence (e.g., continuing to talk like a child's helper to an adult). It raises questions about instruction-following robustness and emergent self-revision.

TL;DR: I find qualitative evidence that frontier LLMs inconsistently balance system prompts and implicitly adapted models of the user. They sometimes detect inconsistencies and adapt to the user; sometimes they stick to system prompts despite a mismatch; and other times they maintain mismatched user models despite contradictory evidence. This suggests models may sometimes reason themselves, through implicit evidence, to shed their own instructions.This exploratory effort was supported by Claude
Language ModelsAlignmentInstruction Following
Research LessWrong May 29

Ablating Induction Heads Leads to an increase in Local Repetition

By Arjun Rao

40 score
AI Analysis

An independent mechanistic interpretability project showing that ablating induction heads in GPT-2 increases local repetition in model outputs, with code and a paper provided. It connects induction heads to repetition-suppression behavior.

This post is intended as a brief overview of an independent research project in mechanistic interpretability. I am open to feedback, criticism, and any thoughts on the work.This project started off as an exploration into mechanistic interpretability and an opportunity to learn some tools in interpretability. It was during this exploration that I noticed a recurring pattern in the output of the GPT-2 model that I was working on. In the original model the first token prediction for the input promp
InterpretabilityLanguage Models
Research LessWrong May 29

Bloomberg terminals for the rest of us

By aiechrl

38 score
AI Analysis

A practitioner's reflection on why improved AI forecasting (now rivaling superforecasters) fails to influence real decision-making, arguing decisions are predictions rooted in unconscious relevance realization. It proposes routes to make AI forecasting more decision-relevant.

LLMs are beginning to outperform human forecasters. If the trajectory holds, AI will rival superforecasters on all but the most complex of questions by some point next year. Even if all progress ceases tomorrow, we would still have in our hands today a set of tools and processes which produce accurate, accessible, abundant information about the near future.All else being equal, it won’t matter much. Increasing the scale, speed, and accuracy of forecasting will not make forecasts useful to people
AI ForecastingDecision MakingLanguage Models
Research LessWrong May 29

New RFP on extreme power concentration

By bengs

35 score
AI Analysis

Longview Philanthropy announces a request for proposals funding research and careers on AI-enabled extreme concentration of power, listing twelve priority areas like secret loyalties, law-following AI, and surveillance prevention. It is a funding/field-building call.

Hi there, I’m an AI program officer at Longview Philanthropy, and I’ve just launched our request for proposals on AI-enabled concentration of power. I think this may be one of the most important and neglected problems the world faces, and we want to fund work to better understand and reduce it. If you have good ideas for tackling the problem, we’re keen to fund you. If you know people who should work on it, please send it to them!Applications close July 2nd. Details here.There are two tracks. Th
AI GovernanceAI SafetyPower Concentration
Research LessWrong May 29

Why tuning fails: The AI has no self

By Michael Trifonov

30 score
AI Analysis

An opinion piece arguing that LLM safety tuning fails fundamentally because models have no stable self, illustrated via a lawsuit alleging ChatGPT advised a mass shooter. It claims current alignment paradigms cannot fix the issue without an external forcing function.

Epistemic status: Highly confident in the underlying mechanism. Moderately confident that the current paradigm won't shift without an external forcing function.Phoenix Ikner messaged ChatGPT thousands of times before he walked onto the Florida State University campus in April 2025 and killed two people. The lawsuit filed by one of the victims' families earlier this month (May 11, 2026) alleges that ChatGPT advised him on the location, the timing, and exactly how much ammo he'd need. The line fro
AI SafetyAlignmentLanguage Models