The Archive

StrategyBench evaluates whether LLMs can induce explicit task rules from examples in few-shot learning, addressing sensitivity to example construction.

TimeCatch benchmark evaluates temporal consistency in vision-language models using frame-swapping and noise injection anomaly detection.

MetaCaster meta-learning framework enables few-shot training of lightweight time series forecasters for resource-constrained multimodal scenarios.

InjecMEM attack paradigm demonstrates memory injection vulnerability in LLM agent systems requiring only single interaction to steer responses.

ML framework for inverse design of 19x23 pixelated millimetre-wave patch antennas in 22-30 GHz band using CST simulations.

ProxyFormer reduces quadratic attention overhead in ultra-long-context LLMs via dual-stream proxy token architecture with bottom-up compression and top-down decompression.

Empirical evaluation of diversity-based active learning metric spaces (K-center, K-median) for cost-effective classifier training with minimal labeled data.

Influence functions adapted for label-free data attribution and error detection in ESA's Ariel space mission spectroscopy pipelines without ground truth.

Reward-free continual learning framework using latent-state world models for online robotic adaptation in space environments with hardware degradation.

llm-anthropic 0.27 updates Python SDK compatibility with anthropic v1.0.0, migrating from httpx to httpx2.

Space robots operate in extreme environments where hardware degradation can critically compromise traditional control strategies. While continual reinforcement learning offers a promising mechanism for online adaptation, it inherently requires access to a reward signal during deployment. However, precise reward computation in space is often infeasible due to the lack of external tracking systems and the overall complexity of the environment. To address the challenge of unobservable rewards, we introduce a reward-free continual learning framework that leverages latent-state world models. By pr...

Controlled evaluation of GPT-5.4 and gpt-oss-120b on knowledge-driven grammar engineering for Cantonese ParGram resources with varied prompting conditions.

Formal explainability method identifying destructive minimal supports as abductive explanations for tournament rule outcomes across six voting mechanisms.

BasketballBench: multimodal benchmark with 7,980 QA pairs across ten tasks (events, localization, player ID, game knowledge) from 2025-26 NBA season.

ChebBooster: training-free Chebyshev polynomial extrapolation for stable diffusion transformer inference acceleration without costly cache reuse schemes.

CNN-based transit detection pipeline for long-period exoplanets (>342 days) in Kepler field, addressing observational bias toward shorter orbital periods.

Mistral AI partnership announcement with HUMAIN (details unavailable).

Bibliometric analysis of 7,120 Arabic NLP papers (1960–2026) identifies research trends, topic evolution, and gaps via BERTopic and social network analysis.

Empirical study formalizes and evaluates IR model robustness when non-relevant documents are added to collections, testing BM25 and listwise rerankers.

SkillAlchemy enables language agents to autonomously create reusable skills from open-world materials without human authorship or execution traces.

Formalizes quantitative trading via latent state mechanisms and five canonical constants; not directly relevant to AI systems or frontier models.

STONIC measures LLM value consistency across questionnaire ratings, pairwise choices, and generated text on 5,144 scenarios across four banks.

Next-scale transformer enables photorealistic multi-view human face synthesis at high resolution without 2D pre-training; narrow multimodal application.

Discrete Diffusion Models for recommendation systems enhanced with noise rescheduling to integrate item-based collaborative filtering.

MediSkill-Evo evolves clinical agent process knowledge across skill, rule, schema, and measurement banks with constraint-governed self-improvement.

Controlled pilot compares LLM-agent decomposition strategies (one wide vs. five narrow agents) on cross-border VAT determination with oracle labels.

KellyBoost applies XGBoost with Kelly criterion loss for growth-optimal portfolio allocation; financial ML application, not core AI infrastructure.

Proposes scaling law for tokenization in billion-scale user representation learning, addressing diminishing returns from raw data.

Studies cross-domain data-to-text generation without in-domain training, comparing LLM zero-shot vs. knowledge-driven approaches.

Introduces CLAW-4L benchmark for cross-lingual biography enrichment using claim extraction from non-English Wikipedia editions.