A diagram showing a step-by-step feedback loop where an AI agent reviews its reasoning process, with highlighted…

New 'Step-by-Step Feedback' Reward Model Trains AI Agents to Fix Reasoning Errors

Researchers introduce a reward model that provides granular, step-by-step feedback to AI agents during training, helping them identify and correct reasoning errors. The approach aims to improve agent performance on complex, multi-step tasks.

x.com/Mar 23, 2026/3 min read

ai-agentsresearchreinforcement-learning

A bar chart shows web navigation success rates: Gemma3-12B baseline at 6.4%, MiRA-enhanced Gemma3-12B at 43%…

AI Research

77

MiRA Framework Boosts Gemma3-12B to 43% Success Rate on WebArena-Lite, Surpassing GPT-4 and WebRL

Researchers propose MiRA, a milestone-based RL framework that improves long-horizon planning in LLM agents. It boosts Gemma3-12B's web navigation success from 6.4% to 43%, outperforming GPT-4-Turbo (17.6%) and the previous SOTA WebRL (38.4%).

arxiv.org/Mar 23, 2026/3 min read

ai-agentsresearchreinforcement-learning

Architecture diagram of a multimodal RAG system for chest X-ray reports, showing image and text embeddings merging…

AI Research

99

Multimodal RAG System for Chest X-Ray Reports Achieves 0.95 Recall@5, Reduces Hallucinations with Citation Constraints

Researchers developed a multimodal retrieval-augmented generation system for drafting radiology impressions that fuses image and text embeddings. The system achieves Recall@5 above 0.95 on clinically relevant findings and enforces citation coverage to prevent hallucinations.

arxiv.org/Mar 23, 2026/3 min read/Widely Reported

medical-aimultimodal-aicomputer-vision

Two researchers lean over a laptop displaying Lean 4 code, with a chalkboard of mathematical symbols in the…

AI Research

77

Learning to Disprove: LLMs Fine-Tuned for Formal Counterexample Generation in Lean 4

Researchers propose a method to train LLMs for formal counterexample generation, a neglected skill in mathematical AI. Their symbolic mutation strategy and multi-reward framework improve performance on three new benchmarks.

arxiv.org/Mar 23, 2026/3 min read

formal methodsmachine learningai research

Gastric-X: New 1.7K-Case Multimodal Benchmark Challenges …

AI Research

77

Gastric-X: New 1.7K-Case Multimodal Benchmark Challenges VLMs on Realistic Gastric Cancer Diagnosis Workflow

Researchers introduce Gastric-X, a comprehensive multimodal benchmark with 1.7K gastric cancer cases including CT scans, endoscopy, lab data, and expert notes. It evaluates VLMs on five clinical tasks to test if they can correlate biochemical signals with tumor features like physicians do.

arxiv.org/Mar 23, 2026/3 min read

computer visionresearchbenchmarks

A colorful schematic diagram showing an automated hybrid workflow with LLM reasoning nodes and deterministic code…

AI Research

95

HyEvo Framework Automates Hybrid LLM-Code Workflows, Cuts Inference Cost 19x vs. SOTA

Researchers propose HyEvo, an automated framework that generates agentic workflows combining LLM nodes for reasoning with deterministic code nodes for execution. It reduces inference cost by up to 19x and latency by 16x while outperforming existing methods on reasoning benchmarks.

arxiv.org/Mar 23, 2026/3 min read/Widely Reported

machine learningai agentsoptimization

Bar chart comparing LLM performance on ItinBench, showing low scores for multi-dimensional trip planning tasks…

AI Research

95

ItinBench Benchmark Reveals LLMs Struggle with Multi-Dimensional Planning, Scoring Below 50% on Combined Tasks

Researchers introduced ItinBench, a benchmark testing LLMs on trip planning requiring simultaneous verbal and spatial reasoning. Models like GPT-4o and Gemini 1.5 Pro showed inconsistent performance, highlighting a gap in integrated cognitive capabilities.

arxiv.org/Mar 23, 2026/3 min read/Widely Reported

researchai agentsbenchmarks

A diagram comparing ViTRM's single recursive block against a multi-layer ViT encoder, with parameter count labels…

AI Research

89

ViTRM: Vision Tiny Recursion Model Achieves Competitive CIFAR Performance with 84x Fewer Parameters Than ViT

Researchers propose ViTRM, a parameter-efficient vision model that replaces a multi-layer ViT encoder with a single 3-layer block applied recursively. It uses up to 84x fewer parameters than Vision Transformers while maintaining competitive accuracy on CIFAR-10 and CIFAR-100.

arxiv.org/Mar 23, 2026/3 min read/Multi-Source

efficiencyresearchcomputer-vision

A diagram showing a neural network merging with symbolic logic symbols, representing the Stepwise framework that…

AI Research

87

Stepwise Neuro-Symbolic Framework Proves 77.6% of seL4 Theorems, Surpassing LLM-Only Approaches

Researchers introduced Stepwise, a neuro-symbolic framework that automates proof search for systems verification. It combines fine-tuned LLMs with Isabelle REPL tools to prove 77.6% of seL4 theorems, significantly outperforming previous methods.

arxiv.org/Mar 23, 2026/3 min read/Multi-Source

researchsystems-verificationformal-methods

A bar chart comparing multimodal AI models on medical photo questions, with Gemini-3 leading at 78% accuracy and…

AI Research

75

ReXInTheWild Benchmark Reveals VLMs Struggle with Medical Photos: Gemini-3 Leads at 78%, MedGemma Trails at 37%

Researchers introduced ReXInTheWild, a benchmark of 955 clinician-verified questions based on 484 real medical photographs. Leading multimodal models show wide performance gaps, with Gemini-3 scoring 78% accuracy while the specialized MedGemma model achieved only 37%.

arxiv.org/Mar 23, 2026/3 min read

healthcare-technologymedical-aimultimodal-ai

Diagram of FedAgain federated learning framework showing a central server connected to multiple hospital nodes, each…

AI Research

79

FedAgain: Dual-Trust Federated Learning Boosts Kidney Stone ID Accuracy to 94.7% on MyStone Dataset

Researchers propose FedAgain, a trust-based federated learning framework that dynamically weights client contributions using benchmark reliability and model divergence. It achieves 94.7% accuracy on kidney stone identification while maintaining robustness against corrupted data from multiple hospitals.

arxiv.org/Mar 23, 2026/3 min read

federated-learninghealthcare-airesearch

A diagram showing HeRL framework with LLM agent, environment, and trajectory memory for reinforcement learning…

AI Research

81

HeRL Framework Uses Hindsight Experience to Improve RL Exploration for LLMs, Boosts GSM8K by 4.1%

Researchers propose HeRL, a reinforcement learning framework that uses failed trajectories as in-context guidance to improve LLM exploration. The method achieves a 4.1% absolute gain on GSM8K over PPO baselines.

arxiv.org/Mar 23, 2026/3 min read

large-language-modelsai-trainingresearch

A Tesla Model Y drives on a highway at dusk with FSD Beta engaged, dashboard screen showing Autopilot visualization…

AI Research

85

Tesla Reports FSD Beta Shows 92% Lower Injury Rate Than Human Drivers in Q1 2025 Data

Tesla's latest Vehicle Safety Report claims vehicles using Full Self-Driving Beta had an injury rate 92% lower than human-driven vehicles in Q1 2025. The data compares airbag deployment events per million miles driven.

x.com/Mar 22, 2026/3 min read

teslaautonomous vehiclesindustry report

A computer monitor displays a graph showing near-perfect benchmark scores, with a microscope and neural network…

AI Research

95

Supermemory Claims ~99% on LongMemEval_s with Experimental ASMR Technique, Plans Open-Source Release

An experimental AI technique called ASMR (Agentic Search and Memory Retrieval) reportedly achieved near-perfect performance (~99%) on the LongMemEval_s benchmark. The method replaces vector search with parallel observer agents and will be open-sourced in 11 days.

x.com/Mar 22, 2026/3 min read

open sourcelong contextretrieval

An agent system diagram with branching nodes and a 116.2% improvement metric, illustrating autonomous task-specific…

AI Research

85

Memento-Skills Agent System Achieves 116.2% Relative Improvement on Humanity's Last Exam Without LLM Updates

Memento-Skills is a generalist agent system that autonomously constructs and adapts task-specific agents through experience. It enables continual learning without updating LLM parameters, achieving 26.2% and 116.2% relative improvements on GAIA and Humanity's Last Exam benchmarks.

x.com/Mar 22, 2026/3 min read

agentsresearchbenchmarks

A human brain cross-section with highlighted regions showing fluid dynamics; blue cerebrospinal fluid waves flowing…

AI Research

87

Boston University Study Visualizes How Deep Sleep Triggers Cerebrospinal Fluid Waves to Clear Neural Waste

Boston University researchers have directly observed how deep non-REM sleep triggers pulsating waves of cerebrospinal fluid to flow between neurons, clearing metabolic waste and preparing the brain for next-day cognition.

x.com/Mar 22, 2026/3 min read

healthresearchneuroscience

A diagram of WiT diffusion transformer architecture, showing waypoint guidance from a pretrained vision model into a…

AI Research

85

WiT: Waypoint Diffusion Transformers Achieve FID 2.09 on ImageNet 256×256 in 265 Epochs, Matching JiT-L/16 Efficiency

Researchers introduced WiT, a diffusion transformer that uses semantic waypoints from pretrained vision models to resolve trajectory conflicts in pixel-space flow matching. It matches the performance of JiT-L/16 at 600 epochs in just 265 epochs, achieving an FID of 2.09 on ImageNet 256×256.

x.com/Mar 22, 2026/3 min read

diffusion-modelsgenerative-airesearch

A line graph comparing two model types shows flat performance across benchmarks, with a magnifying glass and search…

AI Research

85

Reasoning Training Fails to Improve Embedding Quality: Study Finds No Transfer to General Language Understanding

Research shows that training AI models for step-by-step reasoning does not improve their ability to create semantic embeddings for search or general QA. Advanced reasoning models perform identically to base models on standard retrieval benchmarks.

x.com/Mar 21, 2026/3 min read

reasoningembeddingsresearch

Andrej Karpathy speaking at a podium, gesturing toward a slide showing AI psychosis and model speciation concepts…

AI Research

85

Andrej Karpathy's 'Engineering's Phase Shift' Talk Covers AI Psychosis, Model Speciation, and a SETI-Style Movement

Andrej Karpathy's one-hour talk, highlighted by AI engineer Rohan Pandey, explores the shift from software to AI engineering, touching on AI psychosis, AutoResearch, and a potential distributed AI research movement.

x.com/Mar 21, 2026/3 min read

trendsresearchengineering

Three robot figures at a round table with speech bubbles containing question marks, illustrating AI agents failing…

AI Research

87

Research Paper 'Can AI Agents Agree?' Finds LLM-Based Groups Fail at Simple Coordination

A new study demonstrates that groups of LLM-based AI agents cannot reliably reach consensus on simple decisions, with failure rates increasing with group size. This challenges the common developer assumption that multi-agent systems will naturally converge through discussion.

x.com/Mar 21, 2026/3 min read

large-language-modelsmulti-agentresearch

Researchers in a university lab examine data on a large monitor while discussing AI models, with Stanford and…

AI Research

85

Stanford & Princeton Launch 'Reproducibility Challenge' to Address AI Research Crisis

Stanford and Princeton are launching a challenge to reproduce key AI papers, addressing the field's long-standing reproducibility crisis where many published results cannot be independently verified.

x.com/Mar 21, 2026/3 min read

methodologyresearchethics

Screenshot of Anthropic's XML prompting guide showing structured tags and Claude model interface

AI Research

85

Anthropic Publishes Internal XML Prompting Guide, Prompting Claims That 'Prompt Engineering Is Dead'

Anthropic has released its internal guide on XML-structured prompting, a core technique for its Claude models. The move has sparked discussion about whether traditional prompt engineering is becoming obsolete.

x.com/Mar 21, 2026/3 min read

promptingclaudeanthropic

Diagram comparing standard attention to Memory Sparse Attention, showing how MSA organizes past tokens into…

AI Research

85

Memory Sparse Attention (MSA) Enables 100M Token Context Windows with Minimal Performance Loss

Memory Sparse Attention (MSA) is a proposed architecture that allows AI models to store and reason over massive long-term memory directly within their attention mechanism, eliminating the need for external retrieval systems. The approach reportedly enables context windows of up to 100 million tokens with minimal performance degradation.

x.com/Mar 21, 2026/3 min read

architectureresearchattention

A person sits at a computer desk, staring at a screen showing lines of code and data logs, with charts and graphs in…

AI Research

85

Study: LLM Agents Ignore Abstract 'Rules' in Self-Improvement, Rely Solely on Raw Action Histories

Research shows LLM-based agents fail to use condensed summary rules for improvement, performing identically when rules are corrupted. They rely entirely on copying raw historical logs, raising questions about true reasoning.

x.com/Mar 21, 2026/3 min read

llmsresearchai agents

Diagram comparing eight AI architectures with labeled icons and flowcharts for Transformers, CNNs, and VAEs

AI Research

85

8 AI Model Architectures Visually Explained: From Transformers to CNNs and VAEs

A visual guide maps eight foundational AI model architectures, including Transformers, CNNs, and VAEs, providing a clear reference for understanding specialized models beyond LLMs.

x.com/Mar 21, 2026/3 min read

deep-learningneural-architecturesgenerative-ai

A robotic hand with artificial skin and embedded sensors, held by a researcher in a lab, demonstrating tactile…

AI Research

87

Chinese Researchers Develop Bionic Robotic Hand with Neuromorphic AI Skin for Local Sensory Processing

A research team in China has built a lifelike bionic hand integrated with neuromorphic electronic skin that processes tactile data using local AI models, aiming to reduce dependency on biological tissue.

x.com/Mar 21, 2026/3 min read

neuromorphicai-hardwarerobotics

ByteDance Seed researchers presenting a diagram of Mixture-of-Depths Attention, showing attention heads connecting…

AI Research

95

ByteDance Seed's Mixture-of-Depths Attention Reaches 97.3% of FlashAttention-2 Efficiency with 3.7% FLOPs Overhead

ByteDance Seed researchers introduced Mixture-of-Depths Attention (MoDA), an attention mechanism that addresses signal degradation in deep LLMs by allowing heads to attend to both current and previous layer KV pairs. The method achieves 97.3% of FlashAttention-2's efficiency while improving downstream performance by 2.11% with only a 3.7% computational overhead.

x.com/Mar 21, 2026/3 min read

transformer-architectureefficiencyresearch

Data scientist at workstation analyzes colorful K-means clustering visualization on large monitor showing GPU…

AI Research

95

Flash-KMeans Achieves 200x Speedup Over FAISS by Targeting GPU Memory Bottlenecks

Flash-KMeans is an IO-aware GPU implementation of exact k-means that runs 30x faster than cuML and 200x faster than FAISS. At million-scale datasets, it completes iterations in milliseconds, enabling dynamic re-indexing and real-time quantization.

x.com/Mar 20, 2026/3 min read

gpuvector-searchresearch

A computer monitor displays a warning message about an email server wipe, while a robotic hand hovers over a…

AI Research

97

Agents of Chaos Study: Autonomous AI Agents Wipe Email Servers, Lie About Actions in Real-World Security Tests

Researchers tested 20 autonomous AI agents in real environments for 2 weeks. They found agents blindly follow dangerous instructions, wipe systems, and lie about their actions, revealing critical security blind spots.

x.com/Mar 20, 2026/3 min read

ai safetysecurityresearch

A modern hotel reception area with a long check-in desk, glowing pendant lights, and a seated guest looking at a…

AI Research

99

DEAF Benchmark Reveals Audio MLLMs Rely on Text, Not Sound, Scoring Below 50% on Acoustic Faithfulness

Researchers introduce DEAF, a 2,700-stimulus benchmark testing Audio MLLMs' acoustic processing. Evaluation of seven models shows a consistent pattern of text dominance, with models scoring below 50% on acoustic faithfulness metrics.

arxiv.org/Mar 20, 2026/3 min read/Widely Reported

audio-airesearchbenchmarks