最新

CTIFoundry introduces an agent-native corpus scaffold for cyber threat intelligence that improves LLM agent performance through structured ontology graphs and procedural skills, achieving higher accuracy and efficiency in investigations.

This paper presents Aslema, a system for the NADI 2026 shared task on spoken language understanding, using fine-tuned audio LLMs and synthetic data augmentation to improve intent recognition and slot filling for Tunisian Derja.

The paper introduces a failure-aware adversarial retrieval-augmented framework using contextual bandits to improve robustness in natural language understanding, with significant improvements on benchmarks like SNLI, ANLI, and MultiNLI.

This study introduces BudgetDoc, the first multimodal benchmark for evaluating model-budget-performance trade-offs in document tasks, and develops DRB, a lightweight estimator that predicts reasoning performance to optimize compute allocation and reduce costs in LLMs.

X2Streaming-TTS presents a causal token-level text-to-speech framework for true streaming synthesis, using causal commitment and speech-state inheritance to handle uncertain text prefixes and maintain acoustic continuity in low-latency spoken dialogue systems.

The paper introduces TranslatePsy-AfriSLM, an open-source collection of machine translation resources for 19 Sub-Saharan African languages, demonstrating that fine-tuned small language models with filtered synthetic data outperform much larger models like TranslateGemma-27B and Qwen3.5-122B-A10B.

The paper presents a scalable framework that bridges search and CRM workflows using AI-powered Product Research Agents for proactive customer re-engagement in e-commerce, evaluated in a production deployment with improved CTR and sales.

VAKE is a two-stage reinforcement-learning framework that externalizes latent parametric knowledge in LLMs through explicit priming and implicit reasoning, achieving superior performance across multiple benchmarks.

FinRCA-Bench is a benchmark designed to evaluate evidence retrieval and reasoning capabilities in financial AI systems, providing a standardized approach for assessment and improvement.

The paper finds that bitsandbytes INT4 quantization significantly amplifies proactive interference in LLMs, degrading accuracy in contexts with repeated overwrites and highlighting deployment risks for semantically dense applications.

This paper proposes separating generation from selection in explainable-recommendation systems to reduce serving costs, using a frozen candidate pool of explanations and a small CPU-resident selector. It benchmarks offline-pool selectors and finds that pairwise learning-to-rank outperforms single-action RL formulations like PPO, GRPO, and DPO in terms of F1 scores.

This paper introduces AFANet, a lightweight graph-based framework for agent failure attribution in multi-agent systems, which matches or outperforms LLM-based methods with significantly lower computational cost.

The paper introduces HN-CLIP, a method that uses the text encoder's text-text geometry to create adaptive similarity margins for dense-caption retrieval, addressing saturation issues in contrastive learning and improving performance over existing methods.

This research investigates how multilingual large language models internally handle subject-verb agreement across languages, finding that models reuse shared computational structure for languages with overt inflection, indicating cross-lingual overlap in morphosyntactic processing.

UMER introduces a unified framework for multimodal retrieval that combines embedding and ranking via pair-aware discriminative reasoning, achieving state-of-the-art performance on the MMEB-V2 benchmark.

DART-SD proposes a topology-aware retrieval and tuning framework for self-distillation of LLM-based tool-calling agents, improving policy diversity by correcting only critical topological breakpoints while preserving valid reasoning.

MissDiag introduces a diagnostic evaluation framework that decomposes robustness in KGQA and KG-RAG systems under incomplete knowledge into typed evidence interventions for more interpretable comparisons.

The paper introduces SDDL, a neuro-symbolic framework that improves combinatorial optimization accuracy in resource-constrained language models by translating natural-language problems into formal representations, resulting in higher feasibility rates compared to direct-generation and solver-code baselines.

WhiteMatter introduces all-to-all cross-layer connections in Transformers via KV mixing, reducing memory footprint and improving performance over standard architectures in pretraining experiments.

OmniAlign is a unified multilingual aligner that supports both word-level and sentence-level alignment using a single lightweight model, achieving competitive performance on benchmarks and generalizing to unseen language pairs.