arXiv:2506.12382v4 Announce Type: replace-cross Abstract: Ensuring the safety and alignment of Large Language Models is a significant challenge with their growing integration into critical applications and societal functions. While prior research has primarily focused on jailbreak attacks, less attention has been given to non-adversarial failures that subtly emerge during benign interactions. We introduce secondary risks a novel class of failure modes marked by harmful or misleading behaviors during benign prompts. Unlike adversarial attacks, these risks stem from imperfect generalization and often evade standard safety mechanisms. To enable systematic evaluation, we introduce two risk primitives verbose response and speculative advice that capture the core failure patterns. Building on these definitions, we propose SecLens, a black-box, multi-objective search framework that efficiently elicits secondary risk behaviors by optimizing task relevance, risk activation, and linguistic plausibility. To support reproducible evaluation, we release SecRiskBench, a benchmark dataset of 650 prompts covering eight diverse real-world risk categories. Experimental results from extensive evaluations on 16 popular models demonstrate that secondary risks are widespread, transferable across models, and modality independent, emphasizing the urgent need for enhanced safety mechanisms to address benign yet harmful LLM behaviors in real-world deployments.

arXiv:2511.02805v1 Announce Type: cross Abstract: Typical search agents concatenate the entire interaction history into the LLM context, preserving information integrity but producing long, noisy contexts, resulting in high computation and memory costs. In contrast, using only the current turn avoids this overhead but discards essential information. This trade-off limits the scalability of search agents. To address this challenge, we propose MemSearcher, an agent workflow that iteratively maintains a compact memory and combines the current turn with it. At each turn, MemSearcher fuses the user's question with the memory to generate reasoning traces, perform search actions, and update memory to retain only information essential for solving the task. This design stabilizes context length across multi-turn interactions, improving efficiency without sacrificing accuracy. To optimize this workflow, we introduce multi-context GRPO, an end-to-end RL framework that jointly optimize reasoning, search strategies, and memory management of MemSearcher Agents. Specifically, multi-context GRPO samples groups of trajectories under different contexts and propagates trajectory-level advantages across all conversations within them. Trained on the same dataset as Search-R1, MemSearcher achieves significant improvements over strong baselines on seven public benchmarks: +11% on Qwen2.5-3B-Instruct and +12% on Qwen2.5-7B-Instruct relative average gains. Notably, the 3B-based MemSearcher even outperforms 7B-based baselines, demonstrating that striking a balance between information integrity and efficiency yields both higher accuracy and lower computational overhead. The code and models will be publicly available at https://github.com/icip-cas/MemSearcher

arXiv:2510.24701v1 Announce Type: cross Abstract: We present Tongyi DeepResearch, an agentic large language model, which is specifically designed for long-horizon, deep information-seeking research tasks. To incentivize autonomous deep research agency, Tongyi DeepResearch is developed through an end-to-end training framework that combines agentic mid-training and agentic post-training, enabling scalable reasoning and information seeking across complex tasks. We design a highly scalable data synthesis pipeline that is fully automatic, without relying on costly human annotation, and empowers all training stages. By constructing customized environments for each stage, our system enables stable and consistent interactions throughout. Tongyi DeepResearch, featuring 30.5 billion total parameters, with only 3.3 billion activated per token, achieves state-of-the-art performance across a range of agentic deep research benchmarks, including Humanity's Last Exam, BrowseComp, BrowseComp-ZH, WebWalkerQA, xbench-DeepSearch, FRAMES and xbench-DeepSearch-2510. We open-source the model, framework, and complete solutions to empower the community.

When testing an AI model, it’s hard to tell if it is reasoning or just regurgitating answers from its training data. Xbench, a new benchmark developed by the Chinese venture capital firm HSG, or HongShan Capital Group, might help to sidestep that issue. That’s thanks to the way it evaluates models not only on the ability to pass arbitrary tests, like most other benchmarks, but also on the ability to execute real-world tasks, which is more unusual. It will be updated on a regular basis to try to keep it evergreen. 

This week the company is making part of its question set open-source and letting anyone use for free. The team has also released a leaderboard comparing how mainstream AI models stack up when tested on Xbench. (ChatGPT o3 ranked first across all categories, though ByteDance’s Doubao, Gemini 2.5 Pro, and Grok all still did pretty well, as did Claude Sonnet.) 

Development of the benchmark at HongShan began in 2022, following ChatGPT’s breakout success, as an internal tool for assessing which models are worth investing in. Since then, led by partner Gong Yuan, the team has steadily expanded the system, bringing in outside researchers and professionals to help refine it. As the project grew more sophisticated, they decided to release it to the public.

Xbench approached the problem with two different systems. One is similar to traditional benchmarking: an academic test that gauges a model’s aptitude on various subjects. The other is more like a technical interview round for a job, assessing how much real-world economic value a model might deliver.

Xbench’s methods for assessing raw intelligence currently include two components: Xbench-ScienceQA and Xbench-DeepResearch. ScienceQA isn’t a radical departure from existing postgraduate-level STEM benchmarks like GPQA and SuperGPQA. It includes questions spanning fields from biochemistry to orbital mechanics, drafted by graduate students and double-checked by professors. Scoring rewards not only the right answer but also the reasoning chain that leads to it.

DeepResearch, by contrast, focuses on a model’s ability to navigate the Chinese-language web. Ten subject-matter experts created 100 questions in music, history, finance, and literature—questions that can’t just be googled but require significant research to answer. Scoring favors breadth of sources, factual consistency, and a model’s willingness to admit when there isn’t enough data. A question in the publicized collection is “How many Chinese cities in the three northwestern provinces border a foreign country?” (It’s 12, and only 33% of models tested got it right, if you are wondering.)

On the company’s website, the researchers said they want to add more dimensions to the test—for example, aspects like how creative a model is in its problem solving, how collaborative it is when working with other models, and how reliable it is.

The team has committed to updating the test questions once a quarter and to maintain a half-public, half-private data set.

To assess models’ real-world readiness, the team worked with experts to develop tasks modeled on actual workflows, initially in recruitment and marketing. For example, one task asks a model to source five qualified battery engineer candidates and justify each pick. Another asks it to match advertisers with appropriate short-video creators from a pool of over 800 influencers.

The website also teases upcoming categories, including finance, legal, accounting, and design. The question sets for these categories have not yet been open-sourced.

ChatGPT-o3 again ranks first in both of the current professional categories. For recruiting, Perplexity Search and Claude 3.5 Sonnet take second and third place, respectively. For marketing, Claude, Grok, and Gemini all perform well.

“It is really difficult for benchmarks to include things that are so hard to quantify,” says Zihan Zheng, the lead researcher on a new benchmark called LiveCodeBench Pro and a student at NYU. “But Xbench represents a promising start.”

Nature, Published online: 03 July 2024; doi:10.1038/s41586-024-07620-9

Lactylation of NBS1 by TIP60 promotes homologous recombination-driven DNA repair and resistance to chemotherapy in cancer cells and links altered cancer cell metabolism to increase genome stability.

Nature, Published online: 15 January 2024; doi:10.1038/s41586-024-07050-7

Author Correction: A genomic mutational constraint map using variation in 76,156 human genomes

PRESTO-Salsa is a multiplexed GPCR screening platform that enables the interrogation of GPCRome-wide bioactivities across diverse sample types. This platform was used to explore microbiota-induced GPCR activation and revealed the engagement of an immune cell-associated GPCR by a periodontal pathogen protease.

Recurrence represents a well-known poor prognostic factor for colorectal cancer (CRC) patients. This study aimed to establish an effective prognostic prediction model based on noninvasive circulating tumor DNA...

Cell Death Discovery, Published online: 01 November 2022; doi:10.1038/s41420-022-01219-7

Cancer-associated fibroblasts promote the stemness and progression of renal cell carcinoma via exosomal miR-181d-5p