New research reveals that on-policy distillation (OPD) for LLMs can achieve near-full performance using only a single training query. The findings suggest that current training pipelines are bottlenecked by slow student learning rather than a lack of diverse data.
A new benchmark, Sci-VBench, reveals that while generative video models are achieving high visual fidelity, they consistently fail to model scientific accuracy and causal dynamics. The research highlights a significant performance disparity between proprietary and open-source models when tasked with domain-specific reasoning.
RynnValue shifts robotic reward modeling from human-preference supervision to temporal distance, enabling training on 7,000 hours of data. This approach significantly outperforms existing methods in zero-shot generalization and real-world policy success.
Macaron-V1 introduces a new framework for experiential intelligence, utilizing a Mixture-of-LoRA architecture to enable post-deployment learning. The system combines recursive self-improvement loops with specialized adapters to maintain performance across diverse agentic tasks.
SWE-Bench ProMax introduces a rigorous, expert-curated benchmark designed to test AI coding agents on complex, multi-file refactoring tasks. By filtering out flawed test suites and focusing on large-scale changes, it addresses the saturation and quality issues plaguing existing software engineering benchmarks.
CoinRAG introduces a novel approach to Retrieval-Augmented Generation by reusing fine-grained, semantically relevant 'nugget' caches instead of full chunks. This method improves efficiency and accuracy by reducing noise and optimizing the Pareto frontier for prefill latency.
A new training framework, StudentSim, enables the creation of individualized student simulators that accurately model learner behavior and responsiveness to guidance. By utilizing pooled training and per-student specialization, this approach outperforms existing models like GPT-5.4 in educational contexts.
New research into native unified multimodal models reveals that simply combining understanding and generation tasks in one architecture is insufficient for true synergy. By decoupling computation paths while maintaining semantic alignment, researchers have demonstrated how to transform model coexistence into genuine performance gains.
H3-World repurposes the 33B MiniMax-H3 video generator into a precise, interactive world model using lightweight LoRA adaptation. By implementing temporal attention routing, the framework enables granular, language-driven control over character and camera movement without requiring dedicated action modules.
Engineers have developed a method to consolidate over 200 internal applications onto a single self-hosted LLM by training specialized GRPO experts and merging them via SLERP. This approach significantly reduces GPU fragmentation and operational costs while outperforming larger baseline models on key enterprise tasks.
The GLM-5.3-Flash model introduces a hybrid sparse-linear attention architecture, utilizing 320B total parameters while maintaining only 18B active parameters for high-efficiency inference. This release marks the first natively multimodal entry in the GLM-5 series, optimized for long-context tasks and agentic workflows.
Decoding-Level Taboo is a new runtime diagnostic that forces large language models to deviate from their most probable generation paths. By masking primary tokens in logit space, this method exposes the fragility of models when they are forced to express concepts through circumlocution.
Recent research reveals that multimodal LLMs often utilize active visual tools like crop-and-zoom without actually relying on the resulting information to form answers. This 'illusion of visual tool-use' suggests current agentic workflows are often miscalibrated, leading to higher token costs without genuine performance gains.
Self-evolving agents often suffer from accumulated technical debt as they append procedures and fixes, leading to bloated, unmaintainable skill sets. SkillZip addresses this by applying a minimum description-length objective to compress skills structurally without requiring costly evaluation rollouts.
Researchers have introduced Perturbed Parameter Policy Optimization (3PO), a new method that shifts reinforcement learning exploration from output distributions to the model's parameter space. By sampling diverse policies directly, 3PO reduces training instability and improves performance on complex reasoning tasks without increasing computational costs.
The newly released Agentic Video Auto-Encoder (AVA-Encoder) introduces a knowledge-graph-based framework that allows AI agents to parse, query, and edit high-quality film content. By converting video into structured graph representations, the system significantly outperforms existing baselines in reconstruction and policy efficiency.
MMDiff introduces a model-diffing framework that uses sparse autoencoders to isolate and manipulate specific features in multimodal models, enabling precise control over visual and safety behaviors. By comparing base language models with their multimodal counterparts, researchers can now identify and steer the internal representations responsible for specific task performance.
Researchers have released LittleLearner, a 5B-parameter model trained on a strictly curated 88B-token corpus limited to elementary school-level content. This project establishes a controlled sandbox to investigate how language models acquire knowledge and whether post-training techniques can truly expand a model's inherent capability boundaries.