Researchers have introduced RestoreKV and ResKV, two novel methods designed to mitigate the performance degradation inherent in aggressive KV cache compression by reconstructing lost attention information rather than simply discarding tokens.
TurnSight introduces a novel self-distillation framework that improves how LLMs learn to use tools by focusing on turn-level hindsight rather than trajectory-level supervision. By utilizing execution-conditioned hindsight and cross-horizon agreement, the method enables more granular credit assignment in long-horizon reasoning tasks.
Video-DeepResearch introduces a novel framework for multimodal agents that effectively processes continuous video streams through decoupled perception and exploration. By addressing modality bias and parametric knowledge leakage, the model achieves state-of-the-art performance on complex multi-hop VQA tasks.
Researchers have introduced PAST-Bench, a new framework to measure how personal AI agents learn from past experiences over time. The study reveals that while many agents claim to improve, few follow the necessary 'save, retrieve, and update' cycle, leading to the creation of Hermes+ as a more robust alternative.
Researchers have identified a critical numerical underflow issue in ALiBi positional encodings that effectively blinds attention heads at longer sequence lengths. This failure mode, while often masked by standard benchmarks, significantly degrades performance in tasks requiring precise token retrieval.
OPD-V introduces a new paradigm for multimodal on-policy self-distillation that explicitly addresses modality imbalance. By using positive and negative teachers to define a modality-balance trust region, the method improves reasoning performance across diverse MLLM backbones.
The HelloWorld framework introduces a novel self-distillation pipeline and training-free temporal modulation module to enable real-time social interaction in video world models. By allowing users to trigger character responses on demand, it addresses a critical limitation in current generative video systems.
Researchers have successfully adapted NVIDIA's Nemotron retrieval stack for Modern Greek, addressing a critical gap in multilingual RAG capabilities. By fine-tuning a 1B embedder and a 30B-A3B reader, the team achieved significant performance gains, validated by the newly introduced HERA benchmark.
New research reveals that Adam and other coordinate-wise optimizers fail to replicate the implicit low-rank bias found in gradient descent, due to a lack of gauge equivariance. This discovery explains why equivariant optimizers like Muon and Shampoo may offer superior structural learning in matrix factorization tasks.
Researchers have introduced Unsupervised On-Policy Self-Distillation (U-OPSD), a method that allows LLMs to improve reasoning capabilities without relying on external ground-truth signals or environmental feedback. By leveraging internal consistency and majority voting, the approach matches or exceeds supervised distillation methods like GRPO and OPSD across multiple mathematical benchmarks.
MirrorWorld introduces a specialized framework to address the persistent failure of video diffusion models in generating consistent mirror reflections. By decoupling content semantics from geometric alignment, the method enables accurate scene-to-mirror synthesis.
A critical architectural vulnerability in proprietary LLM APIs allows attackers to decrypt and extract reasoning traces by exploiting cross-model compatibility. Researchers have demonstrated that encrypted chain-of-thought blocks can be forced into plaintext by injecting them into less secure models within the same provider's ecosystem.
Scal3R introduces a novel multi-reference pose querying mechanism that decouples local depth estimation from global pose regression. By utilizing lightweight learnable tokens and an online pose-graph optimization system, the method achieves state-of-the-art reconstruction accuracy while mitigating the geometric collapse typical in long-sequence video processing.
The machine translation field faces a crisis of saturation, where standard benchmarks no longer distinguish between model capabilities. The Last Translation Benchmark (LTB) introduces a new paradigm of human-authored, peer-reviewed failure cases designed to break state-of-the-art models and provide actionable evaluation.
LatentStream introduces a 'retrieve-and-internalize' architecture for streaming video understanding, replacing traditional external memory banks with a compact, evolving latent memory. This framework allows Multimodal Large Language Models to maintain long-term context within strict memory budgets.