Researchers introduce Explorative Modeling (XMs), a new paradigm that factors the training loop by exploring K candidate matches between model generations and data, improving performance across images, video, and languag…
Researchers at arXiv (paper 2607.27383v1) established the first convergence guarantees for the plain vector-form Adam optimizer under heavy-tailed stochastic noise, showing it converges to (ρ,ε)-stationary points but wit…
A study by researchers using a YouGov survey of 3,011 U.K. adults found that large language models' alignment with human emotional perception of news framing varies by model, with GPT-5.2 showing the highest correlation …
A new study from arXiv (2607.27384v1) introduces Narrative Anchoring, a failure mode in clinical language models where identical clinical facts expressed in different sociolinguistic registers cause diagnostic outputs to…
Researchers propose BridgeAlign, a preference-alignment pipeline for humanities and social sciences (HSS) tasks, which uses 210k synthetic preference samples to align Qwen3-8B, achieving the best average across 17 benchm…
Researchers introduced HSS-Synth, the first data synthesis pipeline for humanities and social sciences, generating 237,000 instruction-tuning samples that outperform 14 leading baselines on 16 benchmarks. The fine-tuned …
Researchers developed AWARE-FX, an auditable AI/NLP system that converts corporate annual report text into traceable firm-year hedging-disclosure measures, and tested it on 24,909 Hong Kong firm-years from 2008 to 2025, …
Researchers propose Harness-G, a graph-structured retrieval framework that reformulates free-form query generation as finite action selection to address retrieval-equivalence collapse in reinforcement learning search age…
Researchers propose Conditional Retrieval Alignment (CoRA), a gradient-free framework that converts a frozen encoder into a task-conditioned retriever for on-device in-context learning, using closed-form ridge regression…
Researchers present B1ade, an efficient retrieval-augmented generation (RAG) architecture with a 335M parameter embedding model and a 1B parameter small language model (SLM). B1ade-embed achieves top MTEB scores among su…
Researchers propose an unsupervised self-evolving agent framework inspired by diffusion models that enables LLMs to improve specialized skills like creative screenwriting without weight access or external supervision. Th…
Researchers introduced SkillSmith, an augmented large language model that treats model weights as an additional modality, enabling instruction-steered parametric synthesis by ingesting both prefix weights and textual dat…
A new framework called CoevolveSim, introduced by researchers in an arXiv preprint (arXiv:2607.27512v1), shows that introducing finetuned specialist large language models (LLMs) more than doubles the shift in consensus b…
Researchers from an unspecified institution introduced an agentic extraction method using large language models to identify and characterize intertextual reuse in classical Chinese histories, validated against an expert-…
Researchers propose DualAnchor, a gloss-free sign language translation framework that preserves language priors and improves lexical fidelity, achieving strong performance on PHOENIX-2014T and CSL-Daily. The framework us…
Researchers introduced MiGUE-Bench, a benchmark for evaluating large language models (LLMs) on multi-granularity event analysis, covering tasks from event detection to future prediction. The benchmark uses an LLM-driven …
Researchers introduced SparseSpec-L, a training-free self-speculative decoding framework for long-context inference, which achieves up to 2.5x speedup over autoregressive decoding while preserving output distribution. Th…
Google's Agent Development Kit now includes ChronoMem, an open-source semantic version-control layer for LLM agent memory that enables natural-language rollback to prior states. In tests on long-horizon conversational be…
Researchers propose a framework that ensembles the reasoning structure of multiple large language models (LLMs) by weighted merging of Directed Acyclic Graphs (DAGs) extracted from reasoning chains, weighting each step b…
Researchers propose SentiLLM, a framework that uses Semantic-Aligned Structural Abstraction to improve multimodal sentiment analysis by converting non-verbal signals into text-like tokens for large language models. The m…