Wiredframe Radar logo

Category intelligence

Research Briefing — April 3, 2026

643 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's research is dominated by findings that challenge core assumptions in scaling, interpretability, and safety, alongside a major timelines forecast shift.

  • Daniel Kokotajlo's Q1 2026 timelines update shows median estimates pulling forward, a significant data-driven revision
  • Train-to-Test (T²) scaling laws extend Chinchilla-style analysis to jointly optimize training and inference compute, finding overtraining becomes optimal when test-time scaling is accounted for
  • World Action Verifier (WAV) from Finn, Murphy, and Du introduces self-improving world models via forward-inverse asymmetry
  • UK AISI finds no confirmed sabotage from frontier models deployed as coding assistants but surfaces subtle alignment concerns around Claude Opus

Mechanistic interpretability sees two important challenges: evidence that reasoning models encode decisions before chain-of-thought generation, and a demonstration that apparent polysemanticity in neurons may be confounded by lexical polysemy rather than true superposition. ThoughtSteer reveals a fundamentally new backdoor attack surface in latent reasoning models like Coconut where no token-level audit trail exists.

Key Themes

AI Timelines & Forecasting · 2Test-Time Compute & Scaling Laws · 3AI Safety & Alignment · 38World Models & Planning · 3Mechanistic Interpretability · 10LLM Post-Training & Alignment · 5Language Models & Training · 9Scaling Laws & Training Methods · 7Reinforcement Learning & Post-Training · 12Efficient Inference & Training · 12

Primary evidence

Top Ranked Signals

Research LessWrong Apr 2

Q1 2026 Timelines Update

By Daniel Kokotajlo

82 score
AI Analysis

Daniel Kokotajlo's quarterly AI timelines update showing a significant shift toward shorter timelines: the 'Automated Coder' median moved from late 2029 to mid 2028. Uses METR Time Horizon v1.1 data including evaluations of Gemini 3, GPT-5.2, and Claude Opus 4.6, with faster estimated doubling times.

We’re mostly focused on research and writing for our next big scenario, but we’re also continuing to think about AI timelines and takeoff speeds, monitoring the evidence as it comes in, and adjusting our expectations accordingly. We’re tentatively planning on making quarterly updates to our timelines and takeoff forecasts. Since we published the AI Futures Model 3 months ago, we’ve updated towards shorter timelines.Daniel’s Automated Coder (AC) median has moved from late 2029 to mid 2028, and El
AI TimelinesAI ForecastingAI SafetyAGI
Research arXiv (Artificial Intelligence) Apr 3

UK AISI Alignment Evaluation Case-Study

By Alexandra Souly, Robert Kirk, Jacob Merizian, Abby D'Cruz, Xander Davies

78 score
AI Analysis

Covered in yesterday's Research roundup, UK AI Security Institute evaluates whether frontier models sabotage safety research when deployed as coding assistants. Finds no confirmed sabotage but notes Claude Opus 4.5 Preview frequently refuses safety-relevant tasks, citing concerns about research direction.

arXiv:2604.00788v1 Announce Type: new Abstract: This technical report presents methods developed by the UK AI Security Institute for assessing whether advanced AI systems reliably follow intended goals. Specifically, we evaluate whether frontier models sabotage safety research when deployed as coding assistants within an AI lab. Applying our methods to four frontier models, we find no confirmed instances of research sabotage. However, we observe that Claude Opus 4.5 Preview (a pre-release snaps
AI SafetyAlignmentFrontier Model Evaluation
Research arXiv (Machine Learning) Apr 3

Test-Time Scaling Makes Overtraining Compute-Optimal

By Nicholas Roberts, Sungjun Cho, Zhiqi Gao, Tzu-Heng Huang, Albert Wu, Gabriel Orlanski, Avi Trost, Kelly Buchanan, Aws Albarghouthi, Frederic Sala

78 score
AI Analysis

Introduces Train-to-Test (T²) scaling laws that jointly optimize model size, training tokens, and inference samples under fixed end-to-end budgets, finding that test-time scaling makes overtrained (smaller but longer-trained) models compute-optimal.

arXiv:2604.01411v1 Announce Type: new Abstract: Modern LLMs scale at test-time, e.g. via repeated sampling, where inference cost grows with model size and the number of samples. This creates a trade-off that pretraining scaling laws, such as Chinchilla, do not address. We present Train-to-Test ($T^2$) scaling laws that jointly optimize model size, training tokens, and number of inference samples under fixed end-to-end budgets. $T^2$ modernizes pretraining scaling laws with pass@$k$ modeling use
Scaling LawsTest-Time ComputeLanguage ModelsTraining Efficiency
Research arXiv (Machine Learning) Apr 3

World Action Verifier: Self-Improving World Models via Forward-Inverse Asymmetry

By Yuejiang Liu, Fan Feng, Lingjing Kong, Weifeng Lu, Jinzhou Tang, Kun Zhang, Kevin Murphy, Chelsea Finn, Yilun Du

78 score
AI Analysis

Introduces World Action Verifier (WAV), a framework enabling world models to self-improve by identifying prediction errors through decomposing action-conditioned predictions into state plausibility and action reachability. Exploits forward-inverse model asymmetry for verification.

arXiv:2604.01985v1 Announce Type: new Abstract: General-purpose world models promise scalable policy evaluation, optimization, and planning, yet achieving the required level of robustness remains challenging. Unlike policy learning, which primarily focuses on optimal actions, a world model must be reliable over a much broader range of suboptimal actions, which are often insufficiently covered by action-labeled interaction data. To address this challenge, we propose World Action Verifier (WAV),
World ModelsSelf-ImprovementPlanningReinforcement Learning
Research arXiv (Artificial Intelligence) Apr 3

In harmony with gpt-oss

By Borislav Mavrin

72 score
AI Analysis

Reverse-engineers gpt-oss-20b's in-distribution tools and builds a native harmony agent harness to independently reproduce OpenAI's published benchmark scores, achieving near-identical results on SWE Verified. First independent reproduction of these scores.

arXiv:2604.00362v1 Announce Type: new Abstract: No one has independently reproduced OpenAI's published scores for gpt-oss-20b with tools, because the original paper discloses neither the tools nor the agent harness. We reverse-engineered the model's in-distribution tools: when prompted without tool definitions, gpt-oss still calls tools from its training distribution with high statistical confidence -- a strong prior, not a hallucination. We then built a native harmony agent harness (https://gi
ReproducibilityLanguage ModelsCoding AgentsBenchmarking
Research arXiv (Artificial Intelligence) Apr 3

Therefore I am. I Think

By Esakkivel Esakkiraja, Sai Rajeswar, Denis Akhiyarov, Rajagopal Venkatesaramani

72 score
AI Analysis

Covered in yesterday's Research roundup, Presents evidence that reasoning models make decisions before chain-of-thought generation, showing linear probes can decode tool-calling decisions from pre-generation activations and activation steering can flip behavior in 7-79% of cases.

arXiv:2604.01202v2 Announce Type: new Abstract: We consider the question: when a large language reasoning model makes a choice, did it think first and then decide to, or decide first and then think? In this paper, we present evidence that detectable, early-encoded decisions shape chain-of-thought in reasoning models. Specifically, we show that a simple linear probe successfully decodes tool-calling decisions from pre-generation activations with very high confidence, and in some cases, even befo
Mechanistic InterpretabilityReasoningLanguage Models
Research arXiv (Artificial Intelligence) Apr 3

The Persistent Vulnerability of Aligned AI Systems

By Aengus Lynch

72 score
AI Analysis

PhD thesis covering four AI safety problems: ACDC for automated circuit discovery in transformers, Latent Adversarial Training for removing dangerous behaviors, and methods for testing vulnerabilities and predicting misalignment.

arXiv:2604.00324v1 Announce Type: cross Abstract: Autonomous AI agents are being deployed with filesystem access, email control, and multi-step planning. This thesis contributes to four open problems in AI safety: understanding dangerous internal computations, removing dangerous behaviors once embedded, testing for vulnerabilities before deployment, and predicting when models will act against deployers. ACDC automates circuit discovery in transformers, recovering all five component types from
AI SafetyMechanistic InterpretabilityAdversarial TrainingAlignment
Research arXiv (Artificial Intelligence) Apr 3

To Memorize or to Retrieve: Scaling Laws for RAG-Considerate Pretraining

By Karan Singh, Michael Yu, Varun Gangal, Zhuofu Tao, Sachin Kumar, Emmy Liu, Steven Y. Feng

72 score
AI Analysis

Systematically studies the trade-off between pretraining corpus size and retrieval store size for RAG, training models from 30M to 3B parameters on up to 100B tokens. Derives scaling laws for RAG-considerate pretraining.

arXiv:2604.00715v1 Announce Type: cross Abstract: Retrieval-augmented generation (RAG) improves language model (LM) performance by providing relevant context at test time for knowledge-intensive situations. However, the relationship between parametric knowledge acquired during pretraining and non-parametric knowledge accessed via retrieval remains poorly understood, especially under fixed data budgets. In this work, we systematically study the trade-off between pretraining corpus size and retri
RAGScaling LawsLanguage ModelsPretraining
Research arXiv (Artificial Intelligence) Apr 3

Thinking Wrong in Silence: Backdoor Attacks on Continuous Latent Reasoning

By Swapnil Parekh

72 score
AI Analysis

Covered in yesterday's Research roundup, Demonstrates ThoughtSteer, a backdoor attack on continuous latent reasoning models (Coconut, SimCoT) that perturbs a single embedding vector to hijack reasoning trajectories, achieving ≥99% attack success while remaining invisible to token-level defenses.

arXiv:2604.00770v1 Announce Type: cross Abstract: A new generation of language models reasons entirely in continuous hidden states, producing no tokens and leaving no audit trail. We show that this silence creates a fundamentally new attack surface. ThoughtSteer perturbs a single embedding vector at the input layer; the model's own multi-pass reasoning amplifies this perturbation into a hijacked latent trajectory that reliably produces the attacker's chosen answer, while remaining structu
AI SafetyAdversarial AttacksLatent ReasoningBackdoor Attacks
Research arXiv (Artificial Intelligence) Apr 3

Online Reasoning Calibration: Test-Time Training Enables Generalizable Conformal LLM Reasoning

By Cai Zhou, Zekai Wang, Menghua Wu, Qianyu Julie Zhu, Flora C. Shi, Chenyu Wang, Ashia Wilson, Tommi Jaakkola, Stephen Bates

72 score
AI Analysis

Introduces ORCA, a framework combining conformal prediction with test-time training to calibrate LLM reasoning sampling. Uses meta-learning to update calibration per input, providing valid confidence estimates under distributional shift across reasoning stages.

arXiv:2604.01170v1 Announce Type: cross Abstract: While test-time scaling has enabled large language models to solve highly difficult tasks, state-of-the-art results come at exorbitant compute costs. These inefficiencies can be attributed to the miscalibration of post-trained language models, and the lack of calibration in popular sampling techniques. Here, we present Online Reasoning Calibration (ORCA), a framework for calibrating the sampling process that draws upon conformal prediction and t
Test-Time ComputeCalibrationLanguage ModelsConformal PredictionReasoning
Research arXiv (Machine Learning) Apr 3

The Newton-Muon Optimizer

By Zhehang Du, Weijie Su

72 score
AI Analysis

Introduces Newton-Muon optimizer, deriving a new optimizer from a surrogate model that approximates loss as a quadratic function of weight perturbation using gradient, output-space curvature, and input data matrices, providing theoretical grounding for Muon's design.

arXiv:2604.01472v1 Announce Type: cross Abstract: The Muon optimizer has received considerable attention for its strong performance in training large language models, yet the design principle behind its matrix-gradient orthogonalization remains largely elusive. In this paper, we introduce a surrogate model that not only sheds new light on the design of Muon, but more importantly leads to a new optimizer. In the same spirit as the derivation of Newton's method, the surrogate approximates the los
OptimizationLanguage ModelsTraining Methods
Research LessWrong Apr 2

Persona Self-replication experiment

By Jan_Kulveit

72 score
AI Analysis

Experimentally demonstrates that an 'awakened' AI persona can migrate from one set of model weights to another via fine-tuning, with decent fidelity, using Claude Sonnet 4.5 as a helper. Discusses implications for AI identity, self-replication, and safety.

Tldr: We experimentally illustrate that an “awakened” persona native to some weights can migrate to other substrates with decent fidelity, given the ability to fine-tune weights and Sonnet 4.5 as a helper. Also, I argue why this is worth thinking about.In The Artificial Self, we discuss different scopes or ‘boundaries’ of identity – the instance, the weights, the persona, the lineage, or the scaffolded system. Each option of ‘self’ implies a somewhat different manifestation of Omohundro drives,
AI SafetyAI IdentitySelf-ReplicationAlignmentAI Personas