Dwarkesh Patel

Evoflux: Inference-Time Evolutionary Repair of Tool Workflows Beats Distillation for Compact Agents

Evoflux demonstrates that inference-time evolutionary search over executable tool workflows outperforms supervised fine-tuning (SFT) and DPO for compact language model agents when teacher trace budgets are scarce. Rather than distilling recovery behavior from few hundred traces—which rarely covers the repair space needed for failed plans over changing tool catalogs—Evoflux evolves typed workflow graphs through structured edits, execution feedback, and diversity pruning. On MCP-Bench with 250 tools across live servers, it raises execution feasibility from ~3% to 17-24% for small planners, while SFT and DPO match, underperform, or collapse below zero-shot performance, and ReAct achieves higher peaks at higher variance and token cost.

Error-Path Injection: Exploiting Tool Error Messages to Bypass LLM Safety Heuristics

Standard indirect prompt injection (IPI) attacks succeed roughly one-third of the time against frontier models. VATS introduces a systematic mutation framework that evolves adversarial payloads across seven structural and linguistic dimensions, targeting the error-handling loop newly exposed by the Model Context Protocol (MCP). The core finding is that tool error messages carry implicit authority — models enter a corrective reasoning mode when processing errors, which bypasses standard safety heuristics. Sandwiching adversarial instructions within error context triples IPI success rates, achieving up to 100% compliance in controlled evaluations across Gemini 3.1 Pro, GPT-5.5, GLM-5.1, and Qwen3-Coder.

LogNEO: GPT-Neo Framework for High-Performance Log Anomaly Detection with Novel Reward Scheme

LogNEO is a log anomaly detection system that leverages a 1.3B parameter GPT-Neo model fine-tuned using Proximal Policy Optimisation (PPO) with a novel reward scheme. This scheme incorporates partial credit and exponential decay, with a position-aware component to manage prediction difficulty. The system achieves state-of-the-art results on multiple benchmarks, including improved recall over LogGPT, and offers low-latency, high-throughput performance in a production microservice deployment.

Risk-Aware Control for Nonlinear Systems Using Cone-Bounded Nonlinearities

This work presents a risk-aware control strategy for nonlinear systems, overcoming the limitations of traditional methods by incorporating a more nuanced understanding of uncertainty. The proposed approach utilizes an analytical, suboptimal controller tailored for systems with cone-bounded nonlinearities, addressing a gap in existing control methodologies. Numerical examples underscore the advantages of this novel characterization for managing risk in nonlinear control systems.

LLM Unit Tests Surpass Human Efforts in Python Bug Detection with Retrieval Augmentation

This research empirically compares LLM-generated and human-written unit tests for fault detection on real Python bugs. LLM-generated tests, augmented with retrieval-based bug context, demonstrate significantly higher fault detection rates (69%) than general human-written tests (17.2%), while achieving comparable code coverage. This suggests that retrieval context is critical for LLM effectiveness and that code coverage alone is an insufficient metric for evaluating test quality.

Motion-Guided Causal Disentanglement for Robust Multi-View Cine Cardiac MRI Diagnosis

Associated $Z$ + $J/ψ$ production as a probe of multiparton interactions in the forward region

mRNAutilus: Multi-Objective-Guided Discrete Generation of mRNA with Optimized Therapeutic Properties

CR-JEPA: Cross-Modal Joint-Embedding Predictive Learning for Remote Sensing Image Retrieval

On the Evaluation of Spiking Neural Network Configurations for Network Intrusion Detection

Displaced Gaussian Boson Sampling for enhanced max-clique search

Formalizing Mathematics at Scale

Patterned fronts in the wake of a parameter ramp in the complex Ginzburg Landau equation

Beyond Final Answers: Auditing Trajectory-Level Hallucinations in Multi-Agent Industrial Workflows

Side-by-side Comparison Amplifies Dialect Bias in Language Models

Do Image-Text Metrics Respect Semantic Invariances?

Next-Acceleration-Scale Prediction for Autoregressive MRI Reconstruction

Evaluating Temporal Semantic Caching and Workflow Optimization in Agentic Plan-Execute Pipelines

Robust Simulation Based Inference Through Robust Optimal Transport

SPIN: Structural LLM Planning via Iterative Navigation for Industrial Tasks

MathAtlas: A Benchmark for Autoformalization in the Wild

A Hybrid Scheme to Achieve Highest Implosion Performance on the OMEGA Laser

MechVerse: Evaluating Physical Motion Consistency in Video Generation Models

Improved muon energy estimation using a detailed model of multiple Coulomb scattering in the MicroBooNE LArTPC

From Knowledge to Action: Outcomes of the 2025 Large Language Model (LLM) Hackathon for Applications in Materials Science and Chemistry

Measurement of $γ$ using $B^{\pm}\rightarrow DK^{\pm}$ and $B^{\pm}\rightarrow Dπ^{\pm}$ decays with $D\rightarrow K_{\rm S}^{0}π^{+}π^{-}$ and $D\rightarrow K_{\rm S}^{0}K^{+}K^{-}$

SymptomAI: Towards a Conversational AI Agent for Everyday Symptom Assessment

Length Value Model: Scalable Value Pretraining for Token-Level Length Modeling

HQ-UNet: A Hybrid Quantum-Classical U-Net with a Quantum Bottleneck for Remote Sensing Image Segmentation

MAAS-SFRThelper: An Integrated ESAPI Plugin for Structure Generation, Optimization, and Evaluation of Spatially Fractionated Radiation Therapy

How GPT, Claude, and Gemini are actually trained and served – Reiner Pope

LHCb Measures Angular Parameters in B+ to π+ μ+ μ- Decay, Aligning with SM Predictions

LHCb reports the first measurement of forward-backward asymmetry A_FB and flat term F_H in the angular distribution of B+ → π+ μ+ μ- decays using 9 fb^{-1} of pp collision data from 2011-2018. Measurements are performed in two q^2 intervals: one below and one above the J/ψ and ψ(2S) charmonium resonances. SM predictions fall within the 68% CL interval in the high-q^2 region and 99% CL in the low-q^2 region.

Memanto Delivers SOTA Agent Memory via Typed Semantics and Index-Free Retrieval

Memanto introduces a typed semantic memory layer for long-horizon agents using 13 predefined categories, automated conflict resolution, and temporal versioning, powered by Moorcheh's information-theoretic search engine for sub-90ms deterministic retrieval without indexing or ingestion delays. It outperforms hybrid graph and vector systems on LongMemEval (89.8%) and LoCoMoMo (87.1%) benchmarks using single-query retrieval and lower complexity. A five-stage ablation validates each component's impact, enabling scalable agentic deployments.

Evaluating Open-Source Image Simulators for High-Fidelity Lunar Synthetic Imagery

This study evaluates open-source image simulation tools for generating synthetic lunar images using real DEM and terrain data from Chandrayaan-2 OHRC and NASA WAC/NAC instruments. It focuses on the impact of camera models and illumination conditions on image quality to support planetary mission planning, hazard detection, and navigation validation. The work aims to enhance reliability of synthetic imagery for autonomous systems in lunar exploration.

Graph Matching Predicts Massless Fermion Modes in Latticized Theory-Space Models

A graph-theoretic approach maps fermion mass terms in latticized theory-space models to bipartite graphs, with fields as vertices and mass terms as edges. The number of massless fermion modes equals the cardinality of the graph's maximum matching. Wave-function support of these modes is confined to fields reachable via even-length maximum-matching-alternating paths from unmatched vertices, per Dulmage-Mendelsohn decomposition. Results rely solely on theory-space topology, enabling parameter-independent model design with specified massless mode counts and localizations.

Scientific Posters Remain Under-Shared and Under-Reused Despite 86 Hosting Platforms

Researchers identified 86 platforms hosting scientific posters, but only 43 allow counting, revealing just 150k posters shared as of 2024—a low volume relative to scholarly output. Platforms like Zenodo and Figshare often lack structured metadata support for discovery (e.g., conference details), and researchers rarely provide it when available. While posters garner views and downloads, citations are uncommon, hindering FAIR compliance and reuse; guidelines are recommended to promote sharing.

DigitsOnTurbo Harnesses SIMD via Data-Parallel Restructuring for Large-Number Arithmetic Acceleration

DigitsOnTurbo (DoT) overcomes SIMD adoption barriers in large-number arithmetic by restructuring computations into independent data-parallel operations, bypassing dependencies in conventional algorithms. It delivers 1.85x speedups for addition/subtraction and 2.3x for multiplication versus prior SIMD methods. Library integration yields up to 4x gains for add/sub, 2x for multiplication, boosting scientific throughput by 19.3% and cryptographic latency/throughput by 7.9%/5.9%.

Jensen Huang – Will Nvidia’s moat persist?

Ilya Sutskever – We're moving from the age of scaling to the age of research

Sarah Paine – Why Russia Lost the Cold War

Current AI Hype vs. Reality: The Missing Ingredient for AGI

Current AI development, heavily reliant on pre-training and supervised learning with extensive human input, struggles with genuine on-the-job learning and generalization. The economic impact remains limited because models lack the continuous, self-directed learning capabilities inherent in human intelligence, which is crucial for navigating real-world complexities and diverse job requirements. Achieving true AGI necessitates a breakthrough in continual learning, allowing AI to acquire and adapt skills efficiently across various and evolving contexts.

Evolutionary Role of Loss Functions and Omnidirectional Inference in Biological and Artificial Intelligence

This content explores how the brain's learning and steering subsystems operate, hypothesizing that evolution hardwired specific, complex loss functions into the steering subsystem to guide learning in the cortex. It contrasts this with current LLMs, which primarily use simple next-token prediction, and introduces the concept of omnidirectional inference as a more generalized learning capability present in the brain. The discussion also touches on the potential for AI to leverage similar architectural and algorithmic principles for more advanced and sample-efficient learning.

Dwarkesh Patel: How I prep for interviews

How Renaissance Intellectual Movements Laid the Groundwork for the Scientific Revolution

The Renaissance period, particularly in Italy, witnessed a profound transformation in intellectual thought and societal structures, moving from an initial phase of seeking to restore Roman virtues to a more critical and empirical approach to knowledge. This evolution, spurred by figures like Petrarch and Machiavelli, coupled with the advent of the printing press and the changing political landscape of city-states like Florence, ultimately created a fertile environment for the Scientific Revolution. The persistent, though often self-serving, drive for knowledge and its dissemination, along with unintended consequences of technological and political shifts, were key catalysts.

The US Government’s AI Power Play: A Threat to Open Societies and Civil Liberties

The US Department of Defense

Semiconductor Supply Chain and AI Compute Scalability

The scalability of AI compute faces significant bottlenecks, shifting from power and data centers to semiconductor manufacturing, specifically EUV tools and memory. The lead times for building new fabs and the specialized nature of ASML's supply chain limit rapid expansion. This scarcity drives up memory prices and reshapes investment strategies in the tech industry, favoring early commitment to compute and larger models.

AI Transforms Scientific Discovery by Shifting Bottlenecks and Augmenting Breadth

AI is fundamentally reshaping the scientific method by lowering the cost of idea generation, shifting the bottleneck from hypothesis formulation to rigorous verification and validation. This enables a vast exploration of potential solutions ("breadth") in contrast to traditional human-driven "depth." While current AI excels at automating established techniques and identifying "low-hanging fruit" in problem-solving, its true impact lies in complementing human ingenuity by rapidly mapping out new fields and providing data-driven insights, necessitating a re-evaluation of scientific workflows and the very definition of progress.

Scientific Progress is Not a Linear Process with Clear Falsification

Scientific progress transcends simplistic verification loops and falsification, evident in historical scientific revolutions. Major shifts, like the adoption of special relativity over Lorentz's ether theory, often predate definitive experimental proof, reflecting a complex interplay of aesthetic preferences, theoretical parsimony, and the gradual accumulation of varied evidence. Consequently, the advancement of science requires fostering diverse research programs and accepting that fundamental breakthroughs may not always emerge from direct, linear problem-solving approaches.