Tencent's Hyra 2.0 research agent independently creates tools for proteins, mathematical proofs, and chips

Tencent Hunyuan has presented a technical report on Hy-Nomas, its open model for interactive world generation, and released Hy-Nomas-Bench—a test suite for evaluating interactive world video models. The report was posted on arXiv on October 27, following the initial release of the code and weights on October 21.

Hy-Nomas is a world video model based on an 8-billion parameter diffusion core. Users can control it over multiple iterations using free-form text prompts, continuous camera control, or a combination of both methods. The model generates the world segment by segment, ensuring scene consistency over time by conditioning each new segment on the current input and saved history.

According to the report, the key concept involves aligning structured video captions with causal attention that accounts for actions. Descriptive context remains available between segments, while action-related instructions are applied only to the corresponding time window. A vision-and-language-based prompt enhancer transforms user queries and interaction history into structured conditions for the next segment.

The training process follows a multi-stage curriculum. It begins with causal generation, then moves to text interaction and camera control, after which the model trains on histories it generated itself, concluding with a diffusion distillation step. The distilled model requires only four to eight denoising steps. In human evaluations, the team noted that Hy-Nomas surpassed existing world models.

Hy-Nomas-Bench includes 280 test scenarios covering four tasks: multi-turn free-form text interaction, camera navigation, alternating text and camera control, and spatial memory. This suite tests the model's ability to follow sequential instructions, generate physically plausible responses, accurately execute camera movements, and maintain scene consistency when the camera returns to a previously visited location. The kit includes prompts, structured captions, interaction sequences, and reference camera trajectories, and is released under the Apache 2.0 license.

The model itself is also distributed under the Apache 2.0 license. Tencent provides an INT8 configuration that runs on a single RTX 4090 using a sliding window history and CPU offloading, as well as a Diffusers pipeline, LoRA fine-tuning code, and an interactive web demo. Among the released adapters are a cinematic LoRA and a robotics LoRA trained on real data from a two-armed robot.

Developers outside the ecosystem have already integrated support into SGLang, ComfyUI, and WanGP, and one community demo project reported generating 720p in real-time on Google TPUs.

The Qwen team from Alibaba described the process of creating a personal agent, asserting that the next stage in the development of AI assistants depends on the coordinated operation of three components: the model, memory, and service ecosystem. Jiang Guanjun, Head of AI Algorithms Department in Qwen Alibaba's business unit, presented this concept on October 22 at the China National Computer Congress (CNCC) in Chengdu.

Jiang divided the evolution of AI applications into three stages: chat, agent, and assistant. In the chat stage, AI responds to user queries. In the agent stage, the system performs assigned tasks. The assistant, which he says is where the industry is heading, combines professional knowledge with the user's personal context, plans actions around the user's goals, and detects needs they have not noticed. He emphasized that a true personal agent must provide added value and offer proactive, individualized service.

He suggested that theoretically, a personal agent is capable of solving almost all daily problems in work, study, and personal life. However, the main obstacle is the system's ability to correlate its capabilities with the optimal path to solving each specific problem.

Model Development Approach

Regarding the model, Qwen focuses on stable task execution. The team developed what is called a user needs map covering over 30 major industries and detailed it down to more than 8000 atomic requirements for the model's capabilities. Industry experts were involved in evaluating and tuning the model.

Memory Mechanism

Jiang noted that a personal agent must ensure long-term understanding of the user, avoiding the simple accumulation of chat history, which increases noise and computational costs. Qwen constructs its memory based on the degree of event interconnectedness, storing and retrieving them at various levels. For example, an assistant that remembers an elderly user mentioned their pension can give more relevant advice when the same family later asks about finding a job for a child.

Role of the Ecosystem

Concerning the ecosystem, Jiang stated that as models and management methods improve, the ability to link and coordinate external services becomes crucial. Qwen creates a model-oriented open service ecosystem that integrates professional knowledge, tools, industry skills, and third-party agents, while also connecting to e-commerce, local services, healthcare, and education.

In his view, these three elements combine to provide proactive service: memory provides personal information, the model uses expert knowledge to determine needs and plan, and the ecosystem executes the work, such as tracking a patient's rehabilitation schedule or alerting to new financial risks for the user. This report represents the most detailed public description of the technical concept behind the Qwen Go! personal agent, which Alibaba opened for registration this month.

Tongyan Lingzhi, a subsidiary of the Beijing Institute of General Artificial Intelligence's (BIGAI) industrial platform, has introduced its approach to the concept of personal AI. This approach involves creating models that continuously learn based on the data of a specific individual, project, or company, rather than relying solely on queries and information retrieval.

Most existing corporate AI systems today adapt to external users. They use queries to establish rules, extract documents via search, store state in memory, and package reusable steps. However, the model weights themselves rarely change after interaction.

Tongyan Lingzhi asserts that stable habits, such as a team's engineering standards or an investment group's risk management discipline, can be integrated directly into the model parameters. The company's architecture provides for maintaining a general base model for processing common knowledge and reasoning, to which lightweight Personal LoRA adapters are added for each individual user, project, or organization. The company notes that these individual capabilities must be preserved even if the base model is replaced.

To maintain the training of these adapters, the team is developing a mechanism of recursive self-improvement, or RSI. Task traces, human corrections, and resulting outcomes become potential training data. A specialized evaluator, configured according to the standards of each user or company, selects signals worth studying before making an incremental LoRA update. The updated model is then run on the same tasks as the previous version, and only improvements are outputted to the working environment, while regressions are canceled.

CTO Chu Bingbin emphasized that one deviated instance is merely a signal, and only recurring patterns should become training data. Furthermore, the company is working on building infrastructure to make thousands of such adapters available. The TIP system performs training and inference in parallel on the same GPUs, and Paged LoRA manages adapter memory in pages. In one test phase using Qwen2.5-72B-Instruct on eight Nvidia H800 GPUs and an adapter memory budget of 5.79 GB, Paged LoRA demonstrated better latency and tail behavior compared to flat mapping when LoRA queries accounted for 1% to 30% of traffic, although performance decreased at 50%.

Tongyan Lingzhi acknowledges that the work is in its early stages. Although the technical path of RSI and the product architecture are defined, and modules are in the design phase, the fully automated training loop is still under development. Open challenges include catastrophic forgetting, feedback contamination, and model drift.