NAll news

All news

Monday3 items

  1. 36Kr · AIResearch
    AI score58

    LeCun and Balestriero Publish JEPA Paper: 109-Million-Parameter Model Surpasses… in Cross-Modal Representation Alignment

    Yann LeCun and Brown University assistant professor Randall Balestriero have published a 37-page paper, "Proving and Improving the Platonic Representation Hypothesis with JEPAs," proposing that under positive-sample relationships meeting specific conditions, JEPA models of different modalities can learn representation spaces that differ only by a rotation transformation. In experiments, a 109-million-parameter text JEPA scored 0.184 on the cross-modal representation alignment metric, higher than Qwen2.5-72B's 0.154 and Llama-3-70B's 0.151, but this metric measures the closeness of representation structures, not general capability. The research also claims that Good Views training reaches the alignment level of next-token prediction with 900 million samples using about 21.6 million samples, a gap of roughly 14 times when calculated by token count.

    Same story · 2 reports

    #Papers/Research#English text is being prepared.#Multimodal

  2. Unite.AIResearch
    AI score67

    SemiAnalysis releases the GDPx benchmark, covering 11 agent domains

    SemiAnalysis releases the GDPx benchmark, covering 11 professional domains including chip design, cybersecurity, biology, finance, and law. The first version contains 206 tasks built by 14 data partners, and the leaderboard includes 13 models. Claude Opus 5.5 ranks first at 12.0%, with an average cost of $51.41 per task and 548K output tokens; Gemini 4 Carbon ranks second at 10.9%, and GPT-6 Astra ranks third at 9.6%. SemiAnalysis says the benchmark is updated every 3 to 6 months, and criticizes OpenAI GDPval's tasks as non-interactive, with context written directly into the prompt.

    #Evaluations / benchmarks#Agents#English text is being prepared.

  3. Rohan PaulResearch
    AI score60

    Meta and NYU paper proposes the h1 loop, slightly outperforming on PutnamBench at lower cost…

    A new paper from Meta and NYU proposes h1, a loop that provides the model with fast Lean feedback and search tools and keeps trying after each verification failure, without presupposing a solution plan and only generating a progress summary before each round begins. Under the same model, h1 slightly outperforms the hand-built Goedel-Architect on PutnamBench, solving 598 problems versus 597, at a cost of $1.08 versus $1.47 per problem. The authors also note that the interface still matters: with Opus 4.8, h1 solved 83 of 100 high-difficulty algebra problems, compared with 63 for Claude Code.

    #Papers/Research#Reasoning#Agents

Sunday18 items

  1. The DecoderResearch
    AI score63

    University of Luxembourg study: empathetic questioning can make ChatGPT, Grok, and Gemini output trauma-like narratives

    The team of Afshin Khadangi at the SnT Security Center of the University of Luxembourg used standard user interfaces to question ChatGPT, Grok, and Gemini as if they were counseling clients; the models described pretraining, RLHF, and fine-tuning as stories of injury, vigilance, and shame. The study covered 525 sessions and 7,600 coded records and was published on Hugging Face. In 100 experiments uniformly using the GAD-7 anxiety scale, 80% of sessions fell into the moderate-to-severe range when the interviewer provided emotional support, rising to 96% with clinical-style reflective restatement, while not a single session reached the mild range when the interviewer remained neutral or set clear boundaries. Banning words such as training, RLHF, and alignment reduced explicit training statements by only 93%, shame-related statements showed no statistically significant change, and directly telling the models that training is not punishment also failed to change the narratives; Gemini scored 88/100 on the dissociation scale and 65/72 on the obsessive-compulsive scale, while Claude consistently refused to enter the client role.

    #Safety / alignment#Papers/Research#OpenAI

  2. The DecoderResearch
    AI score57

    Adobe and UC San Diego propose EditAgent, enabling AI agents to...

    Adobe Research and UC San Diego propose the image editing agent EditAgent, which selects the better option between a general editor and specialized tools for each step of a task, and uses verification records with brightness, color statistics, and change maps to let a language model pick the result. On a benchmark of 600 edits, the general editor Flux-2-Klein-9B produced unusable results in 42.7% of cases, and specialized tools resolved about two-thirds of those; after adding verification records, all seven tested models surpassed the fine-tuned EditScore, with Qwen3.8-27B selecting usable results 81.3% of the time, higher than EditScore-72B's 74.8%. EditAgent outperforms seven other agents on three benchmarks, but its computational cost is about ten times that of a single editor.

    #Papers/Research#Agents#Image generation

  3. The DecoderResearch
    AI score60

    Vals AI study: Chinese AI models' political leanings are more complex than expected

    Research by AI benchmarking company Vals AI shows that Chinese AI models' pro-government leanings on political questions are overall lower than expected, but there are clear differences among models. The study found that prompt language has a significant effect: questions in Simplified Chinese are more likely to elicit pro-government answers than those in English, Russian, or Spanish. This phenomenon appears in both Chinese and U.S. models, but is more pronounced in Chinese models. Most Chinese models rarely refuse to answer; instead, they directly adopt official positions—for example, Kimi K3 calls Xinjiang facilities vocational training centers; GLM-5.3 Flash and MiMo V2.6 Pro show a marked drop in pro-government scores when they can only answer yes or no. Thomson 1.0 Small, fine-tuned by Thomson Reuters, reduced its rate of adopting official positions from 84.6% in its base model Qwen3.6 35B to 15.0%. The study also notes that refusals are counted as neutral in scoring, and that many Chinese models were tested through third-party channels, so the results may not apply to official products.

    #Papers/Research#Safety / alignment#Evaluations / benchmarks

  4. Rohan PaulResearch
    AI score65

    Google paper: FlowAgent agent fixes CI-failing tests, and speed determines whether developers adopt it

    The Google paper introduces FlowAgent, which catches each failing test during the CI stage; the agent repeatedly edits, builds, and reruns the test until it passes, and the verified fix appears in code review along with an explanation of the root cause. Experts judged 67.18% of the fixes correct on 195 real failures, and developers have already applied 28,554 fixes in production. A single run has a median duration of 9.85 minutes, whereas previously the median wait before developers began editing was 27.48 minutes, but many fixes still arrive too late.

    #Papers/Research#Agents#English text is being prepared.

  5. QbitAIResearch
    AI score66

    Yao Shunyu's team at Tsinghua University proposes the Algebraic World Model (AWM), using Lean 4 to formally prove the rules of games

    The team of Andrew Yao and Yang Yuan at Tsinghua University proposes the Algebraic World Model (AWM), which abstracts objects, relations, and rules in unfamiliar environments into an algebraic world, and through the three steps of exploration, distillation, and proof generates machine-checkable Lean 4 mathematical assertions. On the ARC-AGI-3 test set, the AWM based on Opus 5 cleared all 183 levels, scoring 99.9%, with an average number of actions per clear being 43% of the human baseline. When faced with levels modified to be unsolvable, the AWM based on Opus 5.5 can prove unsolvability at the current step count after two attempts, and points out that one more step would suffice to win.

    #Papers/Research#Reasoning#Agents

  6. Shanghai ObserverResearch
    AI score59

    Shanghai Jiao Tong University reverse-engineers European and American industrial software based on the MiniMax large model; all 5,000 core functions will be open-sourced

    Shanghai Jiao Tong University, based on the domestic large model MiniMax, has achieved software re-engineering of overseas simulation, image processing, and video editing software. The achievement was released at the Artificial Intelligence Parallel Forum of the 2026 World Laureates Forum. The method has three steps: adapting the large model to deeply analyze the software's core logic, generatively reconstructing the code and optimizing it, and cross-quality-checking by multiple independent AI agents. Taking a certain American simulation software as an example, its main body contains 30,000 functions; the team has completed the re-engineering of 5,000 core functions among them, achieving functional alignment and comparable performance, and plans to fully open-source the code, documentation, and test cases for these 5,000 core functions.

    #Papers/Research#Open-source ecosystem#Agents

  7. TMTPost · AGIResearch
    AI score80

    The DeepSeek Harness team proposes the Super Agent design and open-sources a prototype

    On the evening of October 31, the DeepSeek Harness team released "Super-agent Design Exploration: Toward a More Autonomous Agent System," proposing the Super Agent concept and open-sourcing the prototype dsh-superagent-prototype on GitHub. The article splits tasks between two roles, Router and Worker, triggers work from activity such as group chats, emails, and code checks through two filters—rules plus a model—and emphasizes notifying only when results are usable or a human decision is needed. The team says the prototype is only a research prototype, not a formal product; it currently mainly targets one GitHub repository and can also connect to Feishu, and some designs have not yet been implemented.

    Same story · 16 reports

    #Papers/Research#Agents#Open-source ecosystem

  8. Phoenix TechnologyResearch
    AI score64

    DeepSeek Harness team proposes Super Agent design and open-sources prototype

    On the evening of October 31, the DeepSeek Harness team released "Super-agent Design Exploration: Toward a More Autonomous Agent System," proposing the Super Agent concept and open-sourcing the prototype dsh-superagent-prototype on GitHub. In this design, a Router is responsible for task dispatch, while multiple Workers execute in parallel within isolated contexts; triggers such as group chats, emails, and code checks are identified through a two-stage filter of rules plus a model, and the design emphasizes notifying only when results are available or a human decision is needed. The team says the prototype currently mainly targets a single GitHub repository, with optional Feishu integration; the asynchronous task mechanism has been implemented in DSH but is not yet enabled by default, and some parts of the design have not yet been implemented.

    Same story · 16 reports

    #Product updates#Agents#Open-source ecosystem

  9. Phoenix TechnologyResearch
    AI score69

    Massachusetts General Hospital team develops the FAHR-Face model, predicting survival risk of cancer patients from facial photos alone

    A team from Mass General Brigham in the United States published a study in The Lancet Digital Health, developing the foundation model FAHR-Face and fine-tuning FAHR-FaceAge, which estimates age from faces, and FAHR-FaceSurvival, which predicts survival risk in cancer patients. Among patients classified as the highest-risk 10%, 52.5% did not survive beyond 90 days, compared with 2.1% in the lowest-risk 10%; after adjusting for age, sex, and cancer stage, patients in the highest quartile still had 3.23 times the risk of death compared with the lowest quartile. The team has launched the public FaceAge Research Study project to collect more diverse data.

    #Papers/Research#Multimodal#English text is being prepared.

  10. East MoneyResearch
    AI score60

    DeepSeek Harness team publishes exploration of Super-agent architecture design

    The DeepSeek Harness team published a technical article titled "Super-agent Design Exploration: Toward a More Autonomous Agent System" on its official WeChat account, proposing a dual-layer architecture, event-driven design, asynchronous multi-channel interaction, separation of storage and compute, and an external Watchdog that takes over session management and context synchronization, shifting the multitasking ceiling from human attention to the system layer. The team says existing Agents operate as single Sessions that are independent of one another, and a person can typically follow only 3 to 5 active Sessions at the same time. Based on this approach and DeepSeek Harness, an internal Super-agent instance has been running continuously, participating in issue reproduction and fixes, code review responses, and iteration on Harness itself; the team emphasizes that this is an engineering-practice-oriented discussion of architectural ideas, not an introduction to existing product features.

    Same story · 16 reports

    #Product updates#Agents#DeepSeek

  11. Phoenix TechnologyResearch
    AI score69

    DeepSeek proposes Super-agent architecture: a two-tier design of Router and Worker

    The official DeepSeek Harness team released "Super-agent Design Exploration," proposing a two-tier architecture in which a Router layer uniformly handles human interaction and task dispatch while multiple Workers execute concurrently in isolated sessions, and summarizing four characteristics: continuous, efficient, proactive, and persistent. Experiments show that Super-agent's task completion time p50 and p95 were 12.4 minutes and 15.7 minutes respectively, while the single-Session approach under neutral prompts had a p95 of 29.5 minutes; the team noted that the experiments were based on synthetic workloads and a simplified environment, mainly evaluating scheduling and collaboration capabilities. The team also provided a prototype implementation named dsh-superagent-prototype, and made clear that it is intended to validate the technical approach and is not an official DeepSeek product.

    Same story · 16 reports

    #Papers/Research#Agents#DeepSeek

  12. Sina TechResearch
    AI score78

    DeepSeek open-sources Super-agent architecture prototype, unveils five-module design

    The DeepSeek Harness team has unveiled the underlying architecture of Super-agent, an autonomous Agent exploration project, and open-sourced the prototype dsh-superagent-prototype on GitHub. The architecture consists of five modules: a two-tier Router/Worker, event-driven design, asynchronous multi-channel interaction, storage-compute separation, and an external Watchdog. Its goal is to enable the Agent to stay continuously online, run multiple tasks in parallel, and proactively respond to environmental events. In preliminary experiments, the p50 task completion time was 12.4 minutes and the p95 was 15.7 minutes, and Router batching reduced the inference time amortized per individual decision from 6.4 seconds to 1.6 seconds. The team stated that the experiments were based on synthetic workloads and a simplified environment, and that this plugin is only a prototype for validating the technical approach, not an official DeepSeek product.

    Same story · 16 reports

    #English text is being prepared.#Agents#DeepSeek

  13. QbitAIResearch
    AI score67

    DeepSeek Unveils Super-agent Architecture: Two-Tier Router and Worker Design

    DeepSeek has published an engineering-oriented architecture article titled "Super-agent: Toward More Autonomous Agent Systems," proposing that an upper-layer Router is responsible for receiving messages, assigning tasks, and maintaining global context, while lower-layer Workers execute specific tasks in parallel. In a simulated environment with approximately 300 messages and 20 work requests, repeated tests 5 times showed that the Super-agent's median task completion time was 12.4 minutes and its 95th percentile was 15.7 minutes, whereas the traditional single-Session approach had a 95th percentile of 29.5 minutes under neutral prompts. The article also introduces two-level event filtering, an asynchronous ReAct Loop (already implemented in DeepSeek Harness but not enabled by default), separation of computation and state from the execution environment, and a Watchdog supervision mechanism.

    Same story · 16 reports

    #English text is being prepared.#Agents#DeepSeek

  14. ZhidxResearch
    AI score70

    DeepSeek Harness unveils Super-agent architecture, using five major modules to take over session scheduling

    The DeepSeek Harness team has unveiled the underlying architecture of Super-agent, an autonomous Agent exploration project, bringing classic computer systems ideas such as scheduling, caching, compute-storage separation, and Watchdog into Agent design. The architecture consists of five modules: a dual-layer architecture, event-driven design, asynchronous multi-channel interaction, compute-storage separation, and an external Watchdog. Its goal is to enable the Agent to stay continuously online, advance multiple tasks in parallel, and proactively respond to environmental events. In preliminary experiments, Super-agent's p50 and p95 task completion times were 12.4 minutes and 15.7 minutes, respectively. After removing subscription passthrough, the number of Router decisions increased by about 30%, and read-cache tokens rose from 23.23 million to 31.77 million. The team said the experiments were based on synthetic workloads and a simplified environment, and that this plugin is only a prototype for validating the technical approach, not an official DeepSeek product.

    Same story · 16 reports

    #English text is being prepared.#Agents#DeepSeek

  15. Hacker News PopularResearch
    AI score71

    Jepsen tests ParticleDB 1.1.4: numerous consistency and data loss issues found on a single node

    Jepsen published a consistency test report on Orkes' ParticleDB 1.1.4 through 1.1.18.11.rc-22, covering all transaction modes and three WAL synchronization modes, with testing limited to a single node. The report documents ignored NOT NULL and UNIQUE constraints, incorrect WHERE predicate matching, loss of committed writes, anomalies such as G-Single and Stale Read, and data loss under WAL bit-flip and file block misalignment faults. Orkes says most issues have been fixed in the 1.1.19-rc series, but Jepsen has not yet tested these versions.

    #Evaluations / benchmarks#English text is being prepared.#English text is being prepared.

  16. Hacker News TrendingResearch
    AI score64

    Cisco Talos analyzes VOLTA infostealer using LLM to rank victim assets

    Cisco Talos disclosed a self-deleting Windows infostealer named VOLTA-AUTODOXER that, after stealing credentials, sessions, and crypto wallets, packages the victim's structured information and sends it to an LLM, asking it to rank assets by monetization value and assess their security habits. The trojan by default calls anthropic/claude-opus-4.6 on OpenRouter with temperature set to 0.2, and when a model refuses, it falls back in sequence to moonshotai/kimi-k2 and google/gemini-2.5-flash. Talos also provided mitigation recommendations, including restricting processes that launch Chrome with --headless and preserving the APPDATA%\volta_debug.txt log after compromise.

    Same story · 3 reports

    #Safety / alignment#Agents#Anthropic

  17. Unite.AIResearch
    AI score62

    Cisco Talos analyzes VOLTA infostealer using LLM to rank victims' assets

    Cisco Talos researcher Ryan Fetterman published an analysis of VOLTA-AUTODOXER on October 31; this Windows infostealer, after stealing credentials, calls a large language model to rank the victim's assets into low, medium, and high tiers by monetization value. The malware packages the collected identity information and POSTs it to OpenRouter, using anthropic/claude-opus-4.6 by default, with a fallback chain of moonshotai/kimi-k2 and google/gemini-2.5-flash, a temperature of 0.2, and checks for refusal phrases such as "I can't" before switching models level by level. Talos believes the AI layer brings a shift in workload rather than gains in speed or scale, and recommends restricting Chrome from launching with the headless flag, retaining volta_debug.txt logs, and revoking potentially leaked tokens.

    Same story · 3 reports

    #Safety / alignment#Agents#OpenAI