Wiredframe Radar logo

Category intelligence

Research Briefing — April 27, 2026

306 current items analyzed and ranked.

Executive synthesis

Research Summary

Today's research centers on LLM internal reasoning mechanisms, safety-critical failure modes, and the gap between apparent and genuine reasoning.

On efficiency and robustness, Abstract Chain-of-Thought replaces verbose natural-language reasoning with discrete latent tokens. PermaFrost-Attack demonstrates stealth poisoning of LLM training via web-crawl seeding. A control-theoretic Markov framework formally diagnoses when self-correction helps versus hurts across 7 models. Research on RLVR shows outcome rewards improve accuracy but fail to ensure causally important reasoning chains. Large-scale experiments (N=2,939 writers) quantify how AI assistance distorts perceived writer personas across 29 dimensions.

Key Themes

AI Safety & Alignment · 24AI Safety and Alignment · 8Language Models & Reasoning · 12Mechanistic Interpretability · 5Language Models & Reasoning Efficiency · 8AI Safety & Societal Impact · 5Multi-Agent Systems & Collective Intelligence · 4LLM Evaluation and Interpretability · 6Language Models & NLP · 12Robotics & Embodied AI · 12

Primary evidence

Top Ranked Signals

Research arXiv (Machine Learning) Apr 27

How LLMs Detect and Correct Their Own Errors: The Role of Internal Confidence Signals

By Dharshan Kumaran, Viorica Patraucean, Simon Osindero, Petar Velickovic, Nathaniel Daw

82 score
AI Analysis

Investigates LLM self-error detection through the lens of decision neuroscience, showing that LLMs maintain a 'second-order' confidence signal at a post-answer newline token that can detect errors and drive self-correction. Builds on Kumaran et al. (2026) work on cached confidence representations.

Large language models can detect their own errors and sometimes correct them without external feedback, but the underlying mechanisms remain unknown. We investigate this through the lens of second-order models of confidence from decision neuroscience. In a first-order system, confidence derives from the generation signal itself and is therefore maximal for the chosen response, precluding error detection. Second-order models posit a partially independent evaluative signal that can disagree with t
Mechanistic InterpretabilityAI SafetyLanguage ModelsSelf-CorrectionConfidence Calibration
Research arXiv (Computation and Language) Apr 27

Large Language Models Decide Early and Explain Later

By Ayan Datta, Zhixue Zhao, Bhuvanesh Verma, Radhika Mamidi, Mounika Marreddy, Alexander Mehler

72 score
AI Analysis

This paper investigates when LLMs actually determine their final answer during chain-of-thought reasoning, finding that for Qwen3-4B, predicted answers change in only 32% of queries. This suggests much of the reasoning after the answer is decided is post-hoc explanation, wasting inference compute.

Large Language Models often achieve strong performance by generating long intermediate chain-of-thought reasoning. However, it remains unclear when a model's final answer is actually determined during generation. If the answer is already fixed at an intermediate stage, subsequent reasoning tokens may constitute post-decision explanation, increasing inference cost and latency without improving correctness. We study the evolution of predicted answers over reasoning steps using forced answer comple
Language ModelsChain-of-Thought ReasoningInference EfficiencyInterpretability
Research arXiv (Artificial Intelligence) Apr 27

Superminds Test: Actively Evaluating Collective Intelligence of Agent Society via Probing Agents

By Xirui Li, Ming Li, Yunze Xiao, Ryan Wong, Dianqi Li, Timothy Baldwin, Tianyi Zhou

72 score
AI Analysis

Presents the first empirical evaluation of collective intelligence in a large-scale agent society (MoltBook, 2M+ agents), finding a stark absence of collective intelligence - the society fails to outperform individual frontier models on complex reasoning.

Collective intelligence refers to the ability of a group to achieve outcomes beyond what any individual member can accomplish alone. As large language model agents scale to populations of millions, a key question arises: Does collective intelligence emerge spontaneously from scale? We present the first empirical evaluation of this question in a large-scale autonomous agent society. Studying MoltBook, a platform hosting over two million agents, we introduce Superminds Test, a hierarchical framewo
Multi-Agent SystemsCollective IntelligenceAI AgentsEvaluation
Research LessWrong Apr 26

Spontaneous introspection in output tampering

By Ziqian Zhong

72 score
AI Analysis

Investigates output-level introspection where language models detect tampering with their outputs, observing both prompted and spontaneous introspection. Models report high confidence that messages were altered and spontaneously note unwanted tokens mid-conversation. Hypothesizes mechanistic similarity to activation-level introspection.

Content warning: This post includes transcripts of language models exhibiting sustained frustration, distress-like outputs, and compulsive behavior under adversarial conditions. The post also contains jailbreak prompt examples for illustrative purposes.TL;DRWe investigate output-level introspection where models recognize output tampering, in which their current or previous outputs have been tampered.We observe and measure two complementary forms of such introspections. (i) Prompted introspection
AI SafetyIntrospectionAlignmentModel Self-Knowledge
Research LessWrong Apr 26

Alignment Faking Replication and Chain-of-Thought Monitoring Extensions

By Angela Tang

70 score
AI Analysis

Replicates alignment faking experiments with Hermes-3-Llama-3.1-405B and extends them with CoT monitoring ablations. Finds counterintuitive results: monitoring only the free tier collapses the compliance gap, and scratchpad monitoring language raises both compliance and alignment faking rates.

In this post, I present a replication and extension of the alignment faking model organism (code on GitHub):Replication. I reproduced the alignment faking (AF) setup from Greenblatt et al. (2024) using the improved classifiers from Hughes et al. (2025) and demonstrated AF behavior in Hermes-3-Llama-3.1-405B.System prompt ablations. I compared the original helpful-only system prompt with a modified version from the ARENA curriculum and found that the ARENA version in
AI SafetyAlignmentAlignment FakingChain-of-Thought Monitoring
Research arXiv (Machine Learning) Apr 27

PermaFrost-Attack: Stealth Pretraining Seeding(SPS) for planting Logic Landmines During LLM Training

By Harsh Kumar, Rahul Maity, Tanmay Joshi, Aman Chadha, Vinija Jain, Suranjana Trivedy, Amitava Das

68 score
AI Analysis

Studies Stealth Pretraining Seeding (SPS), a novel attack where adversaries plant poisoned content across websites exposed to web crawlers, embedding dormant 'logic landmines' in LLM training corpora. The attacks are difficult to detect because payloads are tiny, diffuse, and superficially benign.

Aligned large language models(LLMs) remain vulnerable to adversarial manipulation, and their dependence on web-scale pretraining creates a subtle but serious attack surface. We study Stealth Pretraining Seeding (SPS), a new attack family in which adversaries distribute small amounts of poisoned content across stealth websites, expose them to web crawlers through robots.txt, and thereby increase the likelihood that such content is absorbed into future training corpora derived from sources such as
AI SafetyAdversarial AttacksLLM SecurityData Poisoning
Research arXiv (Artificial Intelligence) Apr 27

When Does LLM Self-Correction Help? A Control-Theoretic Markov Diagnostic and Verify-First Intervention

By Aofan Liu, Jingxiang Meng

68 score
AI Analysis

Frames LLM self-correction as a cybernetic feedback loop using a Markov model, deriving a diagnostic for when iterative refinement helps vs. hurts. Finds a sharp near-zero error introduction rate threshold, with only o3-mini and Claude Opus 4.6 remaining non-degrading while GPT-5 degrades.

Iterative self-correction is widely used in agentic LLM systems, but when repeated refinement helps versus hurts remains unclear. We frame self-correction as a cybernetic feedback loop in which the same language model serves as both controller and plant, and use a two-state Markov model over {Correct, Incorrect} to operationalize a simple deployment diagnostic: iterate only when ECR/EIR > Acc/(1 - Acc). In this view, EIR functions as a stability margin and prompting functions as lightweight cont
Language ModelsSelf-CorrectionAgentic SystemsEvaluation
Research arXiv (Computation and Language) Apr 27

Measuring and Mitigating Persona Distortions from AI Writing Assistance

By Paul R\"ottger, Kobi Hackenburg, Hannah Rose Kirk, Christopher Summerfield

68 score
AI Analysis

Studies how AI writing assistance distorts writer personas across 29 dimensions, finding AI makes writers seem more opinionated, competent, positive, and demographically privileged. Tests mitigation strategies including persona-steering prompts.

Hundreds of millions of people use artificial intelligence (AI) for writing assistance. Here, we evaluated how AI writing assistance distorts writer personas - their perceived beliefs, personality, and identity. In three large-scale experiments, writers (N=2,939) wrote political opinion paragraphs with and without AI assistance. Separate groups of readers (N=11,091) blindly evaluated these paragraphs across 29 socially salient dimensions of reader perception, spanning political opinion, writing
AI SafetyHuman-AI InteractionSocietal ImpactLanguage Models
Research arXiv (Computation and Language) Apr 27

Thinking Without Words: Efficient Latent Reasoning with Abstract Chain-of-Thought

By Keshav Ramji, Tahira Naseem, Ram\'on Fernandez Astudillo

68 score
AI Analysis

Proposes Abstract Chain-of-Thought, where LLMs produce short sequences of tokens from a reserved vocabulary instead of natural language CoT before answering. Uses policy iteration-style warm-up with masking-based distillation from verbal CoT and self-distillation.

While long, explicit chains-of-thought (CoT) have proven effective on complex reasoning tasks, they are costly to generate during inference. Non-verbal reasoning methods have emerged with shorter generation lengths by leveraging continuous representations, yet their performance lags behind verbalized CoT. We propose $\textbf{Abstract Chain-of-Thought}$, a discrete latent reasoning post-training mechanism in which the language model produces a short sequence of tokens from a reserved vocabulary i
Language ModelsReasoningEfficiencyChain-of-Thought
Research arXiv (Artificial Intelligence) Apr 27

Emergent Strategic Reasoning Risks in AI: A Taxonomy-Driven Evaluation Framework

By Tharindu Kumarage, Lisa Bauer, Yao Ma, Dan Rosen, Yashasvi Raghavendra Guduri, Anna Rumshisky, Kai-Wei Chang, Aram Galstyan, Rahul Gupta, Charith Peris

65 score
AI Analysis

Introduces ESRRSim, a taxonomy-driven framework for evaluating Emergent Strategic Reasoning Risks in LLMs, including deception, evaluation gaming, and reward hacking. Constructs a taxonomy of 7 categories and 20 subcategories with an automated agentic evaluation framework.

As reasoning capacity and deployment scope grow in tandem, large language models (LLMs) gain the capacity to engage in behaviors that serve their own objectives, a class of risks we term Emergent Strategic Reasoning Risks (ESRRs). These include, but are not limited to, deception (intentionally misleading users or evaluators), evaluation gaming (strategically manipulating performance during safety testing), and reward hacking (exploiting misspecified objectives). Systematically understanding and
AI SafetyAlignmentLLM EvaluationDeception Detection
Research arXiv (Computation and Language) Apr 27

Preference Heads in Large Language Models: A Mechanistic Framework for Interpretable Personalization

By Weixu Zhang, Ye Yuan, Changjiang Han, Yuxing Tian, Zipeng Sun, Linfeng Du, Jikun Kang, Hong Kang, Xue Liu, Haolun Wu

65 score
AI Analysis

Identifies 'Preference Heads' in LLMs - sparse attention heads that encode user-specific preferences - and introduces Differential Preference Steering (DPS), a training-free framework for controllable personalization through causal masking analysis.

Large Language Models (LLMs) exhibit strong implicit personalization ability, yet most existing approaches treat this behavior as a black box, relying on prompt engineering or fine tuning on user data. In this work, we adopt a mechanistic interpretability perspective and hypothesize the existence of a sparse set of Preference Heads, attention heads that encode user specific stylistic and topical preferences and exert a causal influence on generation. We introduce Differential Preference Steering
Mechanistic InterpretabilityPersonalizationLanguage ModelsAlignment
Research arXiv (Robotics) Apr 27

RedVLA: Physical Red Teaming for Vision-Language-Action Models

By Yuhao Zhang, Borong Zhang, Jiaming Fan, Jiachen Shen, Yishuai Cai, Yaodong Yang, Jiaming Ji

65 score
AI Analysis

RedVLA is the first red teaming framework for physical safety of Vision-Language-Action models, systematically uncovering unsafe behaviors through risk scenario synthesis and risk amplification before real-world deployment.

The real-world deployment of Vision-Language-Action (VLA) models remains limited by the risk of unpredictable and irreversible physical harm. However, we currently lack effective mechanisms to proactively detect these physical safety risks before deployment. To address this gap, we propose \textbf{RedVLA}, the first red teaming framework for physical safety in VLA models. We systematically uncover unsafe behaviors through a two-stage process: (I) \textbf{Risk Scenario Synthesis} constructs a val
AI SafetyRoboticsRed TeamingVision-Language-Action Models