Articles & Guides

YuE2 combines symbolic composition and audio generation in one pipeline: it first writes a readable melody-and-chord score, then turns that plan into a full song with vocals and accompaniment. Its authors report stronger expert preferences when symbolic planning is enabled.

TraceDance automatically converts undesirable behaviors observed in real agent deployments into targeted evaluation benchmarks. Instead of replaying an entire environment, it tests a model’s next action at recorded decision points against behavior-specific criteria.

A new framework called Physical Coding treats robot state and execution as revisable programs rather than opaque action sequences. Its HexaAnything system uses tools, verification, and feedback to make physical agents more adaptable.

EmbodiedMemory-Bench turns embodied memory into a measurable capability through executable, long-horizon interactions. It evaluates four challenges and introduces a structured external memory system together with the EMem-8B policy.

A KAIST AI study finds that longer self-reflection does not always expand a small reasoning model’s solution space. FlyBy lets the model diagnose its bottleneck and selectively query a stronger model when external knowledge is needed.

A new study finds that a post-trained language model may pass useful capability to another model through single-word answers to unrelated prompts. The authors call the subtle pattern left by post-training a “behavioral shadow.”

A Tsinghua NLP team introduces Diffusion Reward Model, or DRM, which treats reward modeling as conditional density estimation rather than single-value prediction.

Duplex-MPE evaluates speech assistants in shared conversations with several human speakers, asking when they should answer, stay silent, or stop. Its results show that speaking more often is not the same as participating better.

RoboFoundry proposes a “system-as-policy” framework for embodied agents, treating context, memory, skills, and execution interfaces as a unified surface for improvement. The paper reports strong gains across embodied-task, long-horizon memory, and manipulation benchmarks.

Most coding-agent benchmarks still focus on one codebase, while real software changes often span several related repositories. WideSWE introduces 120 real-world tasks to measure whether agents can plan, implement, and verify such coordinated changes.

WorldPlay2 targets a central challenge in interactive world models: delivering real-time control without losing consistency over long rollouts. Its design combines hybrid action-semantic controls, shared memory tokens, and a more stable distillation procedure.

Qwen introduces QwenGyre, an end-to-end framework for online reinforcement learning on extreme-long-horizon agent tasks. It combines elastic GPU allocation with branch-aware trajectory processing and deduplication.

A matched study compares DPO, representation steering, internal probes, and text monitors across safety control and risk detection. Representation methods do not replace behavioral alignment, but they offer useful advantages in low-data and cost-sensitive settings.

A Tencent Hunyuan study compares multimodal language models with and without pretrained visual encoders. Its results suggest that encoder-free models may overcome their early disadvantage as model size and training compute increase.

SkillDRE presents a dual-stage red-team framework for evolving agent skills under both pre-execution scanning and runtime defenses. Its results suggest that evaluating either stage alone can underestimate adaptive skill-based threats.

Moving ML-mediated auction decisions onto privacy-preserving devices also decentralizes budget information. A new audit shows how stale balances can amplify overspending and how payment transformations can create exploitable incentives.

Disaggregated Quantization, or DQ, treats LLM prefill and token-by-token decode as separate systems problems. By specializing formats, weights, and placement for each phase, the approach targets both low-bit accuracy and time-to-first-token performance.

VLA-Precision addresses policy drift and system overhead in real-world online reinforcement learning for vision-language-action models. Its ACoB algorithm combines staged value calibration with reference-regularized updates, while ACoB-Stream targets the throughput bottleneck of large VLAs.

IndicBankBench evaluates language-model assistants in Indian retail banking through 799 cases covering safety, tool use, response quality, and advice. Its results show that occasional success can significantly overstate dependable behavior.

D-JEPA argues that a world model can predict futures accurately while still ranking actions incorrectly. It uses outcomes from executed candidates to reshape decision-critical relations in latent space.

TRACE introduces a condition-aware evaluation framework for streaming video understanding, auditing when evidence becomes valid, how visual history is processed, and when models respond. Its results show that similar QA accuracy can hide major differences in completion, validity, workload, and response timing.