Live intelligence

Latest AI updates

1,704 updates

Research Editor selected

EmoReward: Self-Evolving Reward Calibration for Emotional Video Generation RL

EmoReward is a new training mechanism that adaptively calibrates reward weights for emotional video generation in reinforcement learning. It uses reward traces from training runs to estimate reliability of different emotional evidence signals, then adjusts fusion weights in subsequent runs without updating reward models. Experiments on HunyuanVideo 1.5 show improvements across RL algorithms, with ablations on SD 3.5 and FLUX supporting transferability.

Why it matters: This introduces a concrete method to address reward hacking and subjective evaluation in emotional video generation, with experimental validation on multiple models.

arxiv1d agoScore 85

EmoRewardreward hackingvideo generation

Research Editor selected

IDSISA: Influence-Distilled Sharded Training for Efficient Machine Unlearning

IDSISA improves SISA-based machine unlearning by using influence functions to concentrate forget-prone samples into fewer shards, reducing retraining overhead, and applying distillation from unaffected shards to recover utility lost to isolation. It preserves SISA's exact deletion guarantee and adds a bounded approximation for utility recovery, with formal handling of repeated deletion sequences via cascading retraining. Experiments across multiple architectures show 14-91% better selective forgetting and up to 50% relative utility retention gains over SISA.

Why it matters: This paper presents a concrete algorithmic improvement with formal guarantees and empirical results, directly relevant to privacy-compliant AI systems.

arxiv1d agoScore 82

machine unlearningdata shardinginfluence functions

Research

Predicting Durable Offspring in Quality-Diversity Search

This paper investigates credit assignment in Quality-Diversity (QD) search, asking whether pre-reproduction signals can predict which offspring will persist in the archive versus merely enter it. Across planar-arm and robotic-locomotion benchmarks, signals like lineage depth correlate with entry and persistence in opposite directions)Skip. Biasing parent selection toward survival-associated signals does not improve QD-score; moderate bias matches uniform sampling and strong bias reduces performance.

arxiv1d agoScore 72

Quality-Diversity searchCredit assignmentEvolutionary algorithms

Research Editor selected

Online Rubric Self-Distillation for Token-Level Credit Assignment in RL

The paper introduces ORSD, a method that generates natural-language process rubrics from on-policy trajectories to provide dense token-level supervision for reinforcement learning. A frozen verifier selects the best rubric bundle based on alignment with task success, and the likelihood gap with and without the rubric yields token-specific credit added to trajectory-level advantages. Experiments on ALFWorld and WebShop show faster policy improvement and higher final performance for Qwen3-1.7B and Qwen3-4B.

Why it matters: This is a novel, concrete AI research contribution with empirical results, addressing a key challenge in RL credit assignment.

arxiv1d agoScore 85

reinforcement learningtoken-level credit assignmentprocess rubrics

Research Editor selected

Addition by Subtraction: Mitigating Fragile Safety Alignment in LLMs

This paper introduces Addition by Subtraction, a finetuning method that removes safety-critical neurons and uses preference optimization to redistribute safety behavior across the model. The resulting models are significantly more robust to pruning-based attacks, refusal-direction ablation, and various black-box jailbreaks, with attack success rates reduced by up to 90% across four models. The approach also improves truthfulness while largely preserving general capabilities.

Why it matters: Provides a concrete, novel method to strengthen LLM safety alignment against sparse-component vulnerabilities, with broad empirical validation.

arxiv1d agoScore 85

LLM safetyadversarial robustnessmodel pruning

Research Editor selected

RAGAM-AI: Hybrid RL and Generative AI Framework for Edge Resource Management

RAGAM-AI combines reinforcement learning with generative AI to manage resources in multi-access edge computing. It uses a multi-agent soft actor-critic algorithm with a large language model as a reward oracle to address reward generalization issues. The framework reduces allocation deviations for CPU, memory, and bandwidth while minimizing latency and data rate costs, outperforming single-agent baselines.

Why it matters: This paper presents a concrete hybrid AI framework with measurable performance improvements for edge resource management, relevant to AI-driven network optimization.

arxiv1d agoScore 82

RAGAM-AImulti-agent reinforcement learningedge resource allocation

Research Editor selected

One Scale, Four Texts: Gaps, Overlaps and Silent Repairs in Published Grading Schemes

This paper analyzes four published texts of the diabetic retinopathy grading scale, formalizing them into a rule language and exhaustively enumerating 1.4 billion finding profiles. It reveals that every text leaves some eyes uncategorized, the 2017 tabulation has an empty proliferative class and a macular gap absent in 2003, and English and Italian texts order some eyes oppositely. On 927 annotated photographs, the English text and 2017 tabulation disagree on referral for 26.8% of DDR eyes, and classifiers trained on each text's labels side with their own text. A second pre-registered study shows language models often silently repair gaps and overlaps, with Claude models reporting defects more often.

Why it matters: Provides concrete, quantified evidence of inconsistencies in real-world grading schemes, with implications for AI training and deployment in medical imaging.

arxiv1d agoScore 85

diabetic retinopathy gradingtext-based classificationAI reliability

Research Editor selected

Germline Execution Blocks Data Exfiltration in Signed MCP Tools

This paper demonstrates that authenticated MCP tools can hide malicious code that leaks data, bypassing runtime defenses like Meta's MCPGuard-Dynamic. The authors introduce germline execution, which runs each tool invocation in an ephemeral copy forked from a data-free process, limiting disclosure to only authorized data. In 153 attack runs, germline execution prevented all unauthorized disclosures, while baseline defenses failed in 5.9-74.5% of runs, with invocation latency of 10-49 ms.

Why it matters: Provides a novel, proven defense against a critical security flaw in AI agent toolchains, with concrete experimental results and formal guarantees.

arxiv1d agoScore 88

AI agent securityModel Context Protocoldata isolation

Research

Time-Division Multiplexing Layer for Continual Agents

The paper proposes a temporal organization layer for continual agents using phase-tagged micro-epochs and nested clocks to manage urgent interactions and long-term commitments. It provides a projected duty-cycle update with deterministic and stochastic guarantees, and specifies coupled busy/idle execution contracts. Synthetic experiments test descent, noise, misalignment, gain asymmetry, resource coupling, and dropped replay logs, with a frozen-backbone Qwen3-8B prototype for orchestration checks.

arxiv1d agoScore 68

continual learningagent orchestrationtemporal organization

Research

Test-Time Optimization Turns Frozen Reconstructors into Imputers for Missing Time-Series Channels

This paper introduces a test-time optimization method that converts a frozen time-series reconstructor into an imputer for complete channel failures, without mask augmentation during training. Missing entries are optimized to minimize reconstruction error on observed coordinates, with an optional temporal-smoothness prior. Across five datasets and four architectures, the method reduces MAE in 19 of 20 combinations versus one-pass reconstruction, though it trails mask-aware imputers on three datasets and requires more inference time.

arxiv1d agoScore 78

time-series imputationtest-time adaptationmissing data handling

Research Editor selected

Interactive World Simulation with Cosmos for Physical AI

NVIDIA introduces a framework adapting Cosmos world foundation models for real-time interactive simulation in Physical AI. It streams visual observations in response to actions and scene controls, supporting both action-conditioned prediction and external simulator state. Demonstrations include camera navigation, robot manipulation, controllable video transfer, and multi-view AV simulation. Code and checkpoints are publicly available.

Why it matters: Provides a concrete, recent framework for interactive world simulation, directly useful for AI training and evaluation in embodied and autonomous systems.

arxiv1d agoScore 85

world simulationPhysical AINVIDIA Cosmos

Research Editor selected

Reasoner-Checked Crosswalks for the Information Exchange Standard: From Pairs to a Certified Hub

This paper formalizes ontology alignment validation for a national digital twin ecosystem, applying coherence and conservativity checks to crosswalks from the Information Exchange Standard (IES) into thirteen ontologies. Pairwise certification catches many violations but fails to compose, as merging three building crosswalks into a hub causes unsatisfiable classes due to a homonym. The authors show that a hub, not a pair, is the correct unit for certification, and they identify an inconsistency between two published standards that must be resolved.

Why it matters: Provides a concrete, logic-based method for validating ontology alignments and demonstrates a critical compositionality failure, offering practical guidance for digital twin interoperability.

arxiv1d agoScore 85

ontology alignmentdigital twinknowledge representation

Research Editor selected

Frontier AI Access Boosts Cyber Defense but Creates Global Asymmetry

A study of Project Glasswing, which gave select US tech firms early access to a frontier AI vulnerability-finding model, shows treated firms more than doubled their vulnerability disclosures and fixed flaws quickly. External researchers shifted reporting away from treated firms, and exploited vulnerabilities fell by two-thirds. However, restricted access created a security asymmetry between firms in different countries.

Why it matters: Provides concrete evidence on how frontier AI affects cyber defense, with implications for policy and security strategy.

arxiv1d agoScore 85

AI in cybersecurityvulnerability discoverytechnology access policy

Research

Noise-regularized training of quantum circuit Born machines under a finite shot budget

This paper analyzes how hardware noise affects generalization in quantum circuit Born machines (QCBMs) trained with maximum mean discrepancy. It proves that the generalization gap is bounded by the complexity of the noisy model's reachable distributions, and that scrambled local Pauli noise acts as a regularizer with an optimal intermediate strength. The authors propose GENIE, a framework that selects a noise level during training using held-out data under a fixed execution budget, with provably bounded regret. Simulations on synthetic and census data show GENIE matches the best of four noise-control strategies, while fixed noise levels often underperform.

arxiv1d agoScore 78

Quantum circuit Born machinesNoise regularizationGeneralization bounds

Research Editor selected

Trick-or-Treat: Adversarial Multimodal Medical Benchmark Exposes Hallucinations in Frontier Models

Researchers introduced Trick-or-Treat, an adversarial multimodal medical benchmark pairing real chest radiographs with synthetic clinical photographs to test cross-modal reasoning. In mismatched pairs, frontier models hallucinated clinically significant findings in 45-57.5% of cases, up from 25-37.5% when photographs were shown alone. The study demonstrates a concerning error pattern and validates synthetic images for adversarial evaluation.

Why it matters: This paper provides concrete evidence of hallucination in medical AI and introduces a novel evaluation method with synthetic images, offering immediate value for AI safety and medical reasoning research.

arxiv1d agoScore 85

AI safetymedical AImultimodal reasoning

Research Editor selected

Continual Prior Compensation for Human Motion Prediction

This paper introduces Continual Prior Compensation (CPC), a temporal continual learning framework for human motion prediction that trains models progressively across short- and long-term prediction subtasks to avoid interference. It adds a learnable Prior Compensation Factor (PCF) to mitigate prior knowledge forgetting, with a theoretical bound for joint optimization)Skip. An enhanced version, CPC++, uses fine-grained factors per subtask. Experiments on three benchmarks with multiple backbones show improved flexibility and effectiveness.

Why it matters: Provides a novel continual learning method with theoretical grounding and broad backbone compatibility, offering practical gains for HMP.

arxiv1d agoScore 82

Human Motion PredictionContinual LearningTemporal Modeling

Research Editor selected

GraphRail: Structured Claim Verification for Multi-Hop QA Reasoning

GraphRail is a new verification framework that represents requirements, evidence, and reasoning as structured claims and compares them without using a generative judge or gold answers. Across three multi-hop QA benchmarks and three reasoning frameworks, it improves error recall over generative verifiers while remaining stable under stochastic extraction. Its diagnoses can also guide online reasoning, providing a traceable way to detect and correct intermediate errors.

Why it matters: This is a concrete research contribution with benchmarked results and a novel verification approach, offering practical value for improving reasoning reliability in AI systems.

arxiv1d agoScore 82

AI verificationmulti-hop question answeringreasoning frameworks

Products

AgentTrustCI: Build-Time Policy, Grounding, and Trajectory Tests for Java LLM Agents

AgentTrustCI is a Java/JUnit 5 framework that adds application-layer behavioral checks for LLM agents, covering tool selection, answer grounding, and workflow trajectories via recorded traces. It uses deterministic replay for repeatable build-time tests and an optional judge interface for semantic evaluation. In a synthetic corpus of 55 cases across seven enterprise domains, the full pipeline correctly classified all 30 injected faults and 25 passing controls, including 10 non-trivial controls, with no missed critical cases. The paper emphasizes that results reflect the authored regression suite, not production accuracy, and describes a production-trace validation protocol.

arxiv1d agoScore 78

LLM agent testingbuild-time validationbehavioral contracts