safety

ICLR 2026: 12 papers on making AI systems reliable, efficient, and secure

A 7B agent that beats GPT-4o. Lossless weight compression that speeds up inference by 177%. An arena where 23 teams battled across 103,000 adversarial rounds. This year at ICLR, Lambda is presenting t

safetylambda-labs

All Blog X Post Paper YouTube Reddit GitHub

Clear filters

Safety

I’m in Madrid this week for the first in-person meeting of the United Nations Independent International Scientific Panel on AI. I'm thrilled…

DGX agent

I’m in Madrid this week for the first in-person meeting of the United Nations Independent International Scientific Panel on AI. I'm thrilled to be part of this distinguished group of 40 international

safetyyoshua-bengio--x

23 Apr 2026

Interval POMDP Shielding for Imperfect-Perception Agents

arXiv:2604.20728v1 Announce Type: new Abstract: Autonomous systems that rely on learned perception can make unsafe decisions when sensor readings are misclassified. We study shielding for this setting

Kalman Filter Enhanced GRPO for Reinforcement Learning-Based Language Model Reasoning

arXiv:2505.07527v5 Announce Type: replace Abstract: The advantage function is a central concept in RL that helps reduce variance in policy gradient estimates. For language modeling, Group Relative Pol

Language-Coupled Reinforcement Learning for Multilingual Retrieval-Augmented Generation

arXiv:2601.14896v2 Announce Type: replace Abstract: Multilingual retrieval-augmented generation (MRAG) requires models to effectively acquire and integrate beneficial external knowledge from multiling

Large language models perceive cities through a culturally uneven baseline

arXiv:2604.20048v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly used to describe, evaluate and interpret places, yet it remains unclear whether they do so from a cultural

Learning to count small and clustered objects with application to bacterial colonies

arXiv:2604.20030v1 Announce Type: new Abstract: Automated bacterial colony counting from images is an important technique to obtain data required for the development of vaccines and antibiotics. Howev

Lever: Inference-Time Policy Reuse under Support Constraints

arXiv:2604.20174v1 Announce Type: new Abstract: Reinforcement learning (RL) policies are typically trained for fixed objectives, making reuse difficult when task requirements change. We study inferenc

LLM Agents Predict Social Media Reactions but Do Not Outperform Text Classifiers: Benchmarking Simulation Accuracy Using 120K+ Personas of 1511 Humans

arXiv:2604.19787v1 Announce Type: cross Abstract: Social media platforms mediate how billions form opinions and engage with public discourse. As autonomous AI agents increasingly participate in these

LLM-Guided Safety Agent for Edge Robotics with an ISO-Compliant Perception-Compute-Control Architecture

arXiv:2604.20193v1 Announce Type: new Abstract: Ensuring functional safety in human-robot interaction is challenging because AI perception is inherently probabilistic, whereas industrial standards req

LLMs Can Get 'Brain Rot': A Pilot Study on Twitter/X

arXiv:2510.13928v2 Announce Type: replace-cross Abstract: We propose and test the LLM Brain Rot Hypothesis: continual exposure to junk web text induces lasting cognitive decline in large language mode

MATT-Diff: Multimodal Active Target Tracking by Diffusion Policy

arXiv:2511.11931v2 Announce Type: replace Abstract: This paper proposes MATT-Diff: Multimodal Active Target Tracking by Diffusion Policy, a control policy for active multi-target tracking using a mobi

MedSkillAudit: A Domain-Specific Audit Framework for Medical Research Agent Skills

arXiv:2604.20441v1 Announce Type: new Abstract: Background: Agent skills are increasingly deployed as modular, reusable capability units in AI agent systems. Medical research agent skills require safe

Membership Inference for Contrastive Pre-training Models with Text-only PII Queries

arXiv:2603.14222v2 Announce Type: replace-cross Abstract: Contrastive pretraining models such as CLIP and CLAP, serve as the ubiquitous perceptual backbones for modern multimodal large models, yet the

Memorization, Emergence, and Explaining Reversal Failures: A Controlled Study of Relational Semantics in LLMs

arXiv:2601.02931v2 Announce Type: replace Abstract: Autoregressive LLMs perform well on relational tasks that require linking entities via relational words (e.g., father/son, friend), but it is unclea

MGDA-Decoupled: Geometry-Aware Multi-Objective Optimisation for DPO-based LLM Alignment

arXiv:2604.20685v1 Announce Type: new Abstract: Aligning large language models (LLMs) to desirable human values requires balancing multiple, potentially conflicting objectives such as helpfulness, tru

MOA: Multi-Objective Alignment for Role-Playing Agents

arXiv:2512.09756v2 Announce Type: replace Abstract: Role-playing agents (RPAs) require balancing multiple objectives, such as instruction following, persona consistency, and stylistic fidelity, which

More people will die from suppressing AI than from the imaginary AI apocalypse. They'll die from restricting safe self-driving cars that are…

More people will die from suppressing AI than from the imaginary AI apocalypse. They'll die from restricting safe self-driving cars that are 90% better drivers than people who kill 1.5 million people

safetyyann-lecun--x

Multi-Armed Bandits With Machine Learning-Generated Surrogate Rewards

arXiv:2506.16658v2 Announce Type: replace-cross Abstract: Multi-armed bandit (MAB) is a widely adopted framework for sequential decision-making under uncertainty. Traditional bandit algorithms rely so

Multi-Objective Reinforcement Learning for Generating Covalent Inhibitor Candidates

arXiv:2604.20019v1 Announce Type: new Abstract: Rational design of covalent inhibitors requires simultaneously optimizing multiple properties, such as binding affinity, target selectivity, or electrop

Near-Future Policy Optimization

arXiv:2604.20733v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards (RLVR) has become a core post-training recipe. Introducing suitable off-policy trajectories into on-polic

NeuroSymActive: Differentiable Neural-Symbolic Reasoning with Active Exploration for Knowledge Graph Question Answering

arXiv:2602.15353v2 Announce Type: replace-cross Abstract: Large pretrained language models and neural reasoning systems have advanced many natural language tasks, yet they remain challenged by knowled

Occupancy Reward Shaping: Improving Credit Assignment for Offline Goal-Conditioned Reinforcement Learning

arXiv:2604.20627v1 Announce Type: new Abstract: The temporal lag between actions and their long-term consequences makes credit assignment a challenge when learning goal-directed behaviors from data. G

OnSiteVRU: A High-Resolution Trajectory Dataset for High-Density Vulnerable Road Users

arXiv:2503.23365v2 Announce Type: replace Abstract: With the acceleration of urbanization and the growth of transportation demands, the safety of vulnerable road users (VRUs, such as pedestrians and c

ORPHEAS: A Cross-Lingual Greek-English Embedding Model for Retrieval-Augmented Generation

arXiv:2604.20666v1 Announce Type: cross Abstract: Effective retrieval-augmented generation across bilingual Greek--English applications requires embedding models capable of capturing both domain-speci

Participatory provenance as representational auditing for AI-mediated public consultation

arXiv:2604.20711v1 Announce Type: new Abstract: Artificial intelligence is increasingly deployed to synthesize large-scale public input in policy consultations and participatory processes. Yet no form

Physics-Enhanced Deep Learning for Proactive Thermal Runaway Forecasting in Li-Ion Batteries

arXiv:2604.20175v1 Announce Type: cross Abstract: Accurate prediction of thermal runaway in lithium-ion batteries is essential for ensuring the safety, efficiency, and reliability of modern energy sto

ProMMSearchAgent: A Generalizable Multimodal Search Agent Trained with Process-Oriented Rewards

arXiv:2604.20486v1 Announce Type: new Abstract: Training multimodal agents via reinforcement learning for knowledge-intensive visual reasoning is fundamentally hindered by the extreme sparsity of outc

Recency Biased Causal Attention for Time-series Forecasting

arXiv:2502.06151v2 Announce Type: replace-cross Abstract: Recency bias is a useful inductive prior for sequential modeling: it emphasizes nearby observations and can still allow longer-range dependenc

Relative Principals, Pluralistic Alignment, and the Structural Value Alignment Problem

arXiv:2604.20805v1 Announce Type: cross Abstract: The value alignment problem for artificial intelligence (AI) is often framed as a purely technical or normative challenge, sometimes focused on hypoth

Representational Alignment Across Model Layers and Brain Regions with Multi-Level Optimal Transport

arXiv:2510.01706v2 Announce Type: replace-cross Abstract: Standard representational similarity methods align each layer of a network to its best match in another independently, producing asymmetric re

Resolving space-sharing conflicts in road user interactions through uncertainty reduction: An active inference-based computational model

arXiv:2604.19838v1 Announce Type: new Abstract: Understanding how road users resolve space-sharing conflicts is important both for traffic safety and the safe deployment of autonomous vehicles. While

Rethinking Reinforcement Fine-Tuning in LVLM: Convergence, Reward Decomposition, and Generalization

arXiv:2604.19857v1 Announce Type: cross Abstract: Reinforcement fine-tuning with verifiable rewards (RLVR) has emerged as a powerful paradigm for equipping large vision-language models (LVLMs) with ag

Rodrigues Network for Learning Robot Actions

arXiv:2506.02618v2 Announce Type: replace-cross Abstract: Understanding and predicting articulated actions is important in robot learning. However, common architectures such as MLPs and Transformers l

SAMix: Calibrated and Accurate Continual Learning via Sphere-Adaptive Mixup and Neural Collapse

arXiv:2510.15751v2 Announce Type: replace Abstract: While most continual learning methods focus on mitigating forgetting and improving accuracy, they often overlook the critical aspect of network cali

Sampling-Aware Quantization for Diffusion Models

arXiv:2505.02242v2 Announce Type: replace Abstract: Diffusion models have recently emerged as the dominant approach in visual generation tasks. However, the lengthy denoising chains and the computatio

Semantic Prompting: Agentic Incremental Narrative Refinement through Spatial Semantic Interaction

arXiv:2604.19971v1 Announce Type: cross Abstract: Interactive spatial layouts empower users to synthesize information and organize findings for sensemaking. While Large Language Models (LLMs) can auto

SignDATA: Data Pipeline for Sign Language Translation

arXiv:2604.20357v1 Announce Type: cross Abstract: Sign-language datasets are difficult to preprocess consistently because they vary in annotation schema, clip timing, signer framing, and privacy const

Stochastic Barrier Certificates in the Presence of Dynamic Obstacles

arXiv:2604.20208v1 Announce Type: new Abstract: Safety of stochastic dynamic systems in environments with dynamic obstacles is studied in this paper through the lens of stochastic barrier functions. W

Storm Surge Modeling, Bias Correction, Graph Neural Networks, Graph Convolution Networks

arXiv:2604.20688v1 Announce Type: cross Abstract: Storm surge forecasting remains a critical challenge in mitigating the impacts of tropical cyclones on coastal regions, particularly given recent tren

Temporal Difference Calibration in Sequential Tasks: Application to Vision-Language-Action Models

arXiv:2604.20472v1 Announce Type: cross Abstract: Recent advances in vision-language-action (VLA) models for robotics have highlighted the importance of reliable uncertainty quantification in sequenti

The Existential Theory of Research: Why Discovery Is Hard

arXiv:2604.19810v1 Announce Type: new Abstract: Can scientific discovery be made arbitrarily easy by choosing the right representation, collecting enough data, and deploying sufficiently powerful algo

The Imperfective Paradox in Large Language Models

arXiv:2601.09373v2 Announce Type: replace Abstract: Do Large Language Models (LLMs) genuinely grasp the compositional semantics of events, or do they rely on surface-level probabilistic heuristics? We

The Tool-Overuse Illusion: Why Does LLM Prefer External Tools over Internal Knowledge?

arXiv:2604.19749v1 Announce Type: new Abstract: Equipping LLMs with external tools effectively addresses internal reasoning limitations. However, it introduces a critical yet under-explored phenomenon

Throat and acoustic paired speech dataset for deep learning-based speech enhancement

arXiv:2502.11478v3 Announce Type: replace-cross Abstract: In high-noise environments such as factories, subways, and busy streets, capturing clear speech is challenging. Throat microphones can offer a

Toward Cooperative Driving in Mixed Traffic: An Adaptive Potential Game-Based Approach with Field Test Verification

arXiv:2604.20231v1 Announce Type: new Abstract: Connected autonomous vehicles (CAVs), which represent a significant advancement in autonomous driving technology, have the potential to greatly increase

Understanding Overparametrization in Survival Models through Interpolation

arXiv:2512.12463v3 Announce Type: replace-cross Abstract: Classical statistical learning theory predicts a U-shaped relationship between test loss and model capacity, driven by the bias-variance trade

US child safety group NCMEC received 1.5M reports of suspected CSAM with ties to AI in 2025, a significant surge compared to 67,000 in 2024 and 4,700 in 2023 (Bloomberg)

Bloomberg: US child safety group NCMEC received 1.5M reports of suspected CSAM with ties to AI in 2025, a significant surge compared to 67,000 in 2024 and 4,700 in 2023 — William Michael Haslach was a

safetytechmeme