Instead of releasing Mythos publicly, Anthropic launched Project Glasswing to give a limited group of partners including AWS, Apple, Google, Microsoft, NVIDIA, Cisco, CrowdStrike, JPMorgan Chase, and the Linux Foundation access to the system, backed by $100 million in usage credits and $4 million for open-source security work.

The model reportedly scored 93.9% on SWE-bench Verified and 77.8% on SWE-bench Pro, but its strongest signal came from real-world results, including uncovering a 27-year-old flaw in OpenBSD, a 16-year-old vulnerability in FFmpeg, and autonomously chaining Linux kernel exploits without human input.

Anthropic says Managed Agents is designed to cut the time it takes to move from prototype to production from months to days, with early adopters like Notion, Rakuten, Asana, Vibecode, and Sentry already using it across coding, productivity, and internal workflow automation.

The launch shows Meta is increasingly betting that efficiency, product integration, and distribution, not just model size, will define the next phase of competition in AI.

The model reportedly scored 93.9% on SWE-bench Verified and 77.8% on SWE-bench Pro, but its strongest signal came from real-world results, including uncovering a 27-year-old flaw in OpenBSD, a 16-year-old vulnerability in FFmpeg, and autonomously chaining Linux kernel exploits without human input.

understanding how AI works technically

For example, people who themselves use AI writing tools heavily have been shown to accurately detect AI-written text. A panel of human evaluators can even outperform automated tools in a controlled setting

A study of large-scale web-clicking data employed this theory to explain why certain distributions of web page hits emerge on web sites. Huberman et al. [362] proposed a mathematical model that assumes that at any page, users decide to continue clicking as long as its information scent exceeds some threshold. This information scent can be computed using information foraging theory (IFT).

IFT proposes that information-seeking behavior develops to maximize the rate of information gained per unit of time or effort invested. Note that the term information does not refer to the information-theoretic concept but to subjective interest; here, information means anything that users find interesting.

The roots of IFT lie in optimal foraging theory, originally proposed in biology, which describes the hunting and food search behaviors of animals.

A solution strategy—or

Computational rationality is a theory and a modeling approach rooted in bounded rationality and bounded optimality. Recent applications include typing (Figure 21.7), pointing, driving, multitasking, menu selection, and visual search.

Theories of rationality have increased our understanding of how users fail to be optimal.

MDP is a formalism that originates from studies of sequential decision-making in artificial intelligence and operations research. Instead of the choice between n actions, MDP deals with environments where rewards are delayed (or distal). This requires an ability to plan actions as part of sequences instead of one-shot choices.

Information scent refers to a user's intuition that a cue in the interface represents the information needed. It is an estimation of relevance based on a proximal cue.

IFT proposes that information-seeking behavior develops to maximize the rate of information gained per unit of time or effort invested.

Information foraging refers to information-seeking activities such as navigating, exploring, comparing, searching, or manipulating information contents in an information space.

A payoff refers to the benefits that are left after the costs have been subtracted.

The space of all states, actions, and rewards is called the environment.

Costs are negative rewards that the user incurs for transitioning between states or for being in states that are not good for them.

Visual statistical learning is a research topic in perception that studies how the statistical distribution of our environments affects the deployment of gaze.

bounded rationality states that we are only rational to the extent allowed by the involved constraints, or bounds.

An agent is an actor with the ability to choose actions in pursuit of some goal or reward.

To state that a user's choice is rational means that it is selected with the expectation that it yields the highest utility out of the available options.

It assumes that human long-term memory evolved to help survival by anticipating organismically important events. It is evolutionarily important to remember things that are important for survival. Therefore, the expected value of remembering a thing in the future should affect the probability of recalling it.

According to rational analysis, behavior is sensitive to the statistical distribution of rewards in the environment that a user has experienced. Users learn the way rewards are distributed through continued exposure to an environment and adapt their behavior accordingly. A user's behavior is rational because it is tuned to the distribution of rewards in the environment—the ecology.

The theory assumes that users are 'computationally rational': When picking an action—or deciding how to get from the present state to a state with positive rewards—users are as rational as their cognition allows. Users act based on their often inaccurate and partial beliefs, which they have formed via experience.

Computational rationality is a theory and a modeling approach rooted in bounded rationality and bounded optimality. Recent applications include typing (Figure 21.7), pointing, driving, multitasking, menu selection, and visual search. Its core assumption is that users act in accordance with what they believe is best for them.

Rational analysis is a theory of rational behavior proposed by Anderson and Schooler [21]. It examines the distribution of rewards in the environment to explain how users adapt their behavior. According to rational analysis, behavior is sensitive to the statistical distribution of rewards in the environment that a user has experienced.

They share a focus on the emergence of interactive behavior; in other words, they predict how users choose to behave in certain given circumstances.

Utility refers to the agent's consideration of positive and negative rewards when deciding how to act.

Theories of rationality can be used to inform the design of information environments, addressing considerations such as how to distribute and shape information.

These four theories differ in the factors they include and how the agent's decision-making problem is formulated. As such, the theories differ in how easily they help us find a solution to the user's decision-making problem.

Theories of rationality can make quantitative predictions on user behavior in settings where the user's environment and goals (rewards) are well known.

Theories of rationality do not describe what a user has done but ask what that user could have done. Rational behavior refers to behavior that seeks to maximize the expected utility to the user.

Descriptive theories attempt to capture causes behind behaviour that, from a normative perspective, may appear irrational. This view enables predictions of user behaviour in real-world circumstances.

bounded rationality states that we are only rational to the extent allowed by the involved constraints, or bounds.

The term satisficing is used to describe how users tend to behave when facing a complex decision-making problem. It refers to settling on a satisfactory but not optimal solution in the normative sense.

The author wants to augment the formula to explain the meaning of the terms on either side of the arrow—first

they often benefit from being augmented with descriptive elements, such as labels describing the meaning of an expression or colors linking an identifier to its description in the text.

In this walkthrough, the author is trying to add labels to the formula V(s_t) ← R_t to describe the meaning of its terms in an article they are writing.

Our design was motivated by two major goals for notation authoring. These goals followed from recent studies of notation augmentation [30, 71] and conversations with scientists who had experience writing notation in instructional materials and research communications (4 professors, 2 graduate students, R1–6).

authors sometimes fall back on graphical editors like Google Slides, PowerPoint, Adobe Illustrator, and Mathcha to augment formulas.

Authors of typeset formulas augment those formulas to make them easier to understand.

We define the key projections as markup (in this case, LaTeX), an annotatable render, and a structure hierarchy view. Augmentations are made easy to invoke, and projections are kept synchronized and co-present so that authors can shift between representations as is expedient to them.

the challenge of using these tools is that annotations are unmoored from the structure of the formula and must be redone whenever the formula changes. Authors must perform precision positioning and sizing operations that could be inferred from the coordinates of the augmented expressions.

these markup languages can require cumbersome and error-prone editing, arising from the intermixing of annotation markup with the underlying formula. Participants in a study by Wu et al. [71] identified difficulty with debugging nested braces and locating markup to edit.

lab study participants frequently made errors related to incorrectly matched braces when using a LaTeX baseline to augment formulas.

Authors in Head et al. [30] described that "code gets horrible looking" as macros are added to it to specify augmentations.

FreeForm, a projectional editor wherein authors can augment formulas—with color, labels, spacing, and more—across multiple synchronized representations. Augmentations are created graphically using direct selections and compact menus. Those augmentations propagate to LaTeX markup, which can itself be edited and easily exported.

FreeForm is a projectional editor optimized for notation augmentation. This paper defines the key projections for the text: textual LaTeX, a formula render with tree-aware selections, and a property/hierarchy view.

The key ideas are that authors can interact with different projections of the formula at will, and graphical edits are ultimately connected back to the formula via markup.

Authors of typeset formulas augment those formulas to make them easier to understand.

Ply offers this LLM-supported program decomposition supported by visualization and parameterization UIs, permitting users to use interactions beyond chat to compose their programs incrementally.

designing complex behavior can be a difficult programming task, and program representations in end-user programming tools may not be well-suited for heavy programs.

Ply allows users to develop, test, and tweak program components, exploring possibilities for how data can be transformed and composed to discover and achieve goals. This style of programming can support many use cases, even those not traditionally considered in the trigger-action programming model.

Through the combination of these features, Ply allows users to develop, test, and tweak program components, exploring possibilities for how data can be transformed and composed to discover and achieve goals.

Frequently, code-generation systems focus on building and then refining a full working application, using visibility of the full underlying code as a fallback when users need to build understanding of the generated program.

the simplicity of links between triggers and actions limits the expressivity of such systems.

Ply provides users with tools to build components incrementally, creating new layers on top of existing components that "wrap" the behavior of underlying layers.

When building a linkage, Ply identifies parameters of the implementation that may be tweaked to customize the behavior of the linkage.

Each sensor is accompanied by a glanceable visualization of the sensor's output payloads on the Ply canvas. This visualization is specific to the sensor and its output type, showing the most critical information for evaluating whether the sensor is behaving as expected.

Ply uses a server program written in TypeScript to make code generation requests to a large language model and to execute the resulting code, which passes messages to and from sensors and actuators.

Each layer in Ply tracks its dependencies; sensors receive data from their dependencies, actuators push data to their dependencies, and linkages each refer to exactly one sensor and one actuator dependency. Collections of layers and linkages in Ply are isomorphic to node graphs in node-based programming languages.

Code generation offered by large language models can serve to author this glue code for trigger-action programs, allowing for data from triggers to be mapped to input data for actions automatically even when their native data formats or intended functionality do not match exactly.

Ply allows users to develop, test, and tweak program components, exploring possibilities for how data can be transformed and composed to discover and achieve goals. This style of programming can support many use cases, even those not traditionally considered in the trigger-action programming model.

users can develop, test, and tweak program components, exploring possibilities for how data can be transformed and composed to discover and achieve goals.

It encourages program decomposition into "layer" abstractions, It automatically creates visualizations of event payloads at layer boundaries to help users understand layer behavior without having to read the underlying generated code, and It constructs ad hoc parametrization interfaces that allow users to configure important dimensions of the behavior of each layer without having to re-author it.

Ply maintains the simplicity of a straightforward connection between a trigger and action but provides a structure within which users can enlist an LLM to specify the behavior of each trigger and action.

However, such LLM-authored code, especially when implementing nontrivial logic, can be difficult to specify, understand or debug. Users need appropriate tools and handles to understand and make changes to the computation that is being performed in such code.

Trigger-action programming offers an elegant interface to construct simple programs that result in customized behavior for software or devices.

Trigger-action programming has been a success in end-user programming. Traditionally, the simplicity of links between triggers and actions limits the expressivity of such systems. LLM-based code generation promises to enable users to specify more complex behavior in natural language. However, users need appropriate ways to understand and control this added expressive power.

We also discuss the role of AI in science, including AI safety.

external evaluations of the passing paper also uncovered hallucinations, faked results, and overestimated novelty

one manuscript achieved high enough scores to exceed the average human acceptance threshold, marking the first instance of a fully AI-generated paper successfully navigating a peer review.

Qwen3.5 397B A17B: 15.3%, DeepSeek V3.2: 14.5%, GLM-5: 14.5%, Kimi K2.5: 11.5%, MiniMax-M2.7: 10.6%

GPT-5.4 (xhigh) scores the highest on APEX-Agents-AA Pass@1 with a score of 33.3%, followed by Claude Opus 4.6 (Adaptive Reasoning, Max Effort) with a score of 33.0%, and Gemini 3.1 Pro Preview with a score of 32.0%

Context is basically how many things a machine can keep in its operational memory - it's not so different from the very human cognitive load.

their productivity is affected by the state of the codebase.

Code is upstream of all other applications because it's the core building block for any piece of software, so AI's accelerating impact on code should accelerate every other domain.

accounting and auditing showing nearly a 20 percent jump on GDPval and even domains like police / detective work showing a nearly 30 percent improvement.

Jack Cheng considers Pip, his Plus One, somewhere between a colleague and pet with a personality—one he programmed himself, drawing on references from Studio Ghibli, bird watching, and Catherine O'Hara.

70 percent refer to their Plus Ones by gendered pronouns.

We're writing the etiquette in real time.

A "parallel organization chart," in which each AI worker has a name, manager, and job description, allows your company to move faster than it ever could with humans alone.

In UTAUT, Venkatesh extended TAM by incorporating two constructs not directly related to a system's perceived properties, but derived from external aspects: social influence and facilitating conditions. Additionally, UTAUT posits four mediating factors that moderate the impact of each key construct on usage intention and behavior, namely gender, age, experience, and voluntariness of use.

While our key focus is to build a theoretical model that explains the process through which older adults accept (or reject) mobile technology, which can provide theoretical guidelines when designing a technology, and which may also be able to generate new investigations and experiments.

We analyzed the second-round interview data using inductive and deductive approaches informed by grounded theory and other qualitative analysis methods [33, 22].

We inductively analyzed the first-round interview data using thematic analysis based on a grounded theory approach [33]. Grounded theory methods build theory iteratively from the data, using rigorous coding practices.

Technology acceptance has been widely studied, and several models have been proposed and tested [10, 37]. However, the HCI literature lacks a comprehensive explanation of technology acceptance among older adults.

Azjen's theory of planned behavior [1, 2] posits that a specific behavior is the result of an intention to carry it out, and that intention is determined by attitudes, norms, and the perception of control over the behavior. Drawing upon this theory of planned behavior, Davis et al. developed the technology acceptance model (TAM) [10].

To summarize, existing models of technology acceptance can provide a partial explanation of older adults' behaviors of mobile technology acceptance. However, we also identified critical elements that are not represented in the existing models. Components in red boldface in Figure 3 provide a preview of the new elements we have identified and their relationship to the components proposed in earlier models.

by triangulating our empirical findings with existing theoretical models from the literature, we found out that the existing models of technology adoption require new theory components to be able to describe technology adoption processes of our participants. In particular, we identified an additional phase that is prominent among the participants, intention to learn, but did not appear in prior models. Then, we identified three new factors that significantly influence their technology acceptance but which are, again, not represented in the existing models: self-efficacy, conversion readiness, and peer support.

we found out that the existing models of technology adoption require new theory components to be able to describe technology adoption processes of our participants. In particular, we identified an additional phase that is prominent among the participants, intention to learn, but did not appear in prior models. Then, we identified three new factors that significantly influence their technology acceptance but which are, again, not represented in the existing models: self-efficacy, conversion readiness, and peer support.

Then, by triangulating our empirical findings with existing theoretical models from the literature, we found out that the existing models of technology adoption require new theory components to be able to describe technology adoption processes of our participants.

We identified three distinct factors that influence older adults' technology acceptance behaviors, particularly the intention to learn phase, that are not represented in prior models: self-efficacy, conversion readiness, and peer support.

The human's job is to curate sources, direct the analysis, ask good questions, and think about what it all means. The LLM's job is everything else.

it almost always traces back to the interface rather than the language model

since reasoning models and agentic AI can rack up quite a bill

Fields that are not exposed now will become exposed in the future

Exposure alone is a completely meaningless tool for predicting displacement

Raising prices will for sure decrease demand and that risks killing the growth story. And even if revenue keeps growing, it doesn’t matter if there are no margins

they don’t have to spend it to win. It’s a defensive move for them, if they commit $50B, OpenAI and Anthropic need to go raise $100B each to stay competitive

That’s up 20x in six weeks. This idea, called tokenmaxxing, is the deliberate practice of maximizing token consumption.

The platform doesn’t need to bother with individual prompts - it just needs to see where the questions cluster.

They meet their target S-parameter specifications despite having very alien-looking geometries.

coding agents are themselves becoming formidable instruments of attack

select known-vulnerable dependency versions 50% more often than humans.

the entities making dependency decisions are increasingly not human.

harness combinations doesn't shrink as models improve. Instead, it moves

There's an old saying that content is king. With agents, context is.

You don't need a separate agent API. You need to look at every `input()` call, every CWD assumption, every pretty-printed-only output, and ask: what if the user on the other end is a process, not a person?

Implicit state is the Enemy

The funny part is that none of this made the CLI worse for humans. The TUI picker still works and looks fancy, progress spinners still spin, confirmation dialogs still confirm. We just added a second door.

Every prompt is a flag in disguise

Designing for agents forced us to build better tools for everyone.

If ChatGPT was the moment consumers discovered AI could talk, OpenClaw may be the moment they discovered AI could act.

As AI moves from a destination to a feature, our methodology will need to shift.

纯粹收集分析这种形态，过去互联网有过先例，但你会发现它卖不出去钱。

AIサイエンティストは、アイデアの創出から実験、分析、論文執筆、そして査読に至るまでの科学的研究サイクル全体をAIが自律的に遂行する仕組みです。この仕組みの定量的評価も含めた結果を、共同研究者とともにNature誌の論文として公開しています。

By late next year, the rate of model releases and the number of new evals required could be such that even keeping ourselves informed will be a challenge without effective AI assistance.

Some recent models that don't currently have time horizons: Gemini 3.1 Pro, GPT-5.2-Codex, Grok 4.1

AI agents are typically several times faster than humans on tasks they complete successfully.

Case study: blackmail

Functional emotions may work quite differently from human emotions, and do not imply that LLMs have any subjective experience of emotions, but appear to be important for understanding the model's behavior.

We refer to this phenomenon as the LLM exhibiting functional emotions: patterns of expression and behavior modeled after humans under the influence of an emotion, which are mediated by underlying abstract representations of emotion concepts.

Our key finding is that these representations causally influence the LLM's outputs, including Claude's preferences and its rate of exhibiting misaligned behaviors such as reward hacking, blackmail, and sycophancy.

Create multilingual experiences that go beyond translation and understand cultural context.

E2B and E4B · Try in Google AI Edge Gallery

Gemma 4 models undergo the same rigorous infrastructure security protocols as our proprietary models.

Build autonomous agents that plan, navigate apps, and complete tasks on your behalf, with native support for function calling.

E2B & E4B · A new level of intelligence for mobile and IoT devices

frontier AI companies can run more of the best AIs to speed up their own AI research, relative to their competitors. Right now these gains are maybe noticeable but not game-changing, but that'll probably change in the next few years.

Tang Jie (CEO of Zhipu AI) even recently said: "The truth may be that the gap [between US and Chinese AI] is actually widening."

American hyperscalers are driving a data center buildout that's larger than the Manhattan Project and Apollo Program at their peaks.

Just last year, Anthropic spent over ten times more on compute than Minimax and Zhipu AI combined, and the gap is even wider for OpenAI:

We estimate Google is the largest single owner of AI compute, holding about one quarter of global cumulative capacity as of Q4 2025.

Because these benchmarks are human-authored, they can only test for risks we have already conceptualized and learned to measure.

From anthropic.com

New Anthropic Fellows Research: a new method for surfacing behavioral differences between AI models.

Cloud agents produce demos and screenshots of their work for you to verify.

We're introducing Cursor 3, a unified workspace for building software with agents.

In the last year, we moved from manually editing files to working with agents that write most of our code.

With Uni-1, we are laying the foundation for a system that can see, speak, reason, and imagine in one continuous stream.

This unified design naturally extends beyond static images to video, voice agents, and fully interactive world simulators.

We evaluate on ODinW-13 following consistent protocols from prior work. ODinW (Open Detection in the Wild) measures open vocabulary dense detection, testing fine-grained visual reasoning.

Uni-1 shows that learning to generate images materially improves fine-grained visual understanding performance, reasoning over regions, objects, and layouts.

Uni-1 can perform structured internal reasoning before and during image synthesis. It decomposes instructions, resolves constraints, and plans composition, then renders accordingly.

Uni-1 is a multimodal reasoning model that can generate pixels.

Reference-guided generation with source-grounded controls.

Common-sense scene completion, spatial reasoning, and plausibility-driven transformation.

Built on Unified Intelligence, Uni-1 understands intention, responds to direction, and thinks with you.

Uni-1 ranks first in human preference Elo for Overall, Style & Editing, and Reference-Based Generation, and second in Text-to-Image.

New AI models, especially those from Anthropic,have triggered a new set of actions for how we build and secure our products.

AI-powered analysis uncovers data at a scale and depth that legacy frameworks were not designed to accommodate.

including Anthropic’s latest unreleased AI model–Claude Mythos Preview.

it also lowers the threshold for attackers, empowering less-skilled actors to launch complex, high-impact campaigns.

We do not plan to make Claude Mythos Preview generally available, but our eventual goal is to enable our users to safely deploy Mythos-class models at scale.

AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities.

Cai et al. [117] interviewed 21 pathologists who used a deep neural network to aid in thediagnosis of prostate cancer. The interviews showed that pathologists needed to learn moreabout the network’s strengths and limitations to use it effectively. They also wanted to knowthe design objective of the network and the kind of data on which it was trained.

The performance of the system must be reliable and controllable. Its behavior should be safe, and the way it is designed and used should be ethical [768]. Users need to trust the system's decisions and ability. It should be made clear to the user what it can and cannot do.

such systems should be designed to take into account the fact that automated results will inevitably be incorrect on occasion.

users can be trained to understand not only the decision-making tasks but also the underpinning capabilities and limitations of the automation solution.

automated systems that indicate when automation may fail or has failed are more likely to gain an appropriate level of trust from users.

When the system fails, users need to be able to redirect it. To avoid biases and discrimination, some level of transparency and explainability is required.

For small entrepreneurs in the US, deciding what to sell and where to make it has traditionally been a slow, labor-intensive process that can take months. Now that work is increasingly being done by AI tools like Accio, which help connect businesses with manufacturers in countries including China and India.

Zhang, of Alibaba.com, says Accio currently does not include advertising. Suppliers can pay for higher placement in Alibaba.com's regular search results, but Zhang says Accio is 'not integrated' with that system.

Sellers say that while AI tools have made it easier to come up with ideas and get a business off the ground, they do not replace the core skills that make someone good at e-commerce.

Sally Li, a representative at a makeup packaging company in Wuhan, China, says her firm has started writing more detailed product descriptions and adding information about its equipment and manufacturing experience on Alibaba.com because it suspects those details make its listings more likely to be surfaced by AI.

McClary took the process from there, contacting the supplier himself to discuss the revised design. Within a month, the new version of the Guardian flashlight was back up for sale on Amazon and on his brand's website.

current approaches often rely on decoupled trigger-response pipelines or are limited to captioning-style narration, reducing their effectiveness for open-ended question answering and long-horizon interaction

amplifies the false narrative that technology and creativity are at odds, and that existing rights holders must be compensated by AI companies for changing industry dynamics.

The government has so far favoured a pro-innovation, sector-led approach, prioritising voluntary principles over hard regulation.

introducing a commercial text and data mining exception for AI training would expand the AI sector in the country.

This article argues that squirrel ecology offers a sharp comparative case because arboreal locomotion, scatter-hoarding, and audience-sensitive caching couple all three demands in one organism.

Agentic AI is increasingly judged not by fluent output alone but by whether it can act, remember, and verify under partial observability, delay, and strategic observation.

We've seen customers go from 10-20% field accuracy with a frontier model to 99-100% just by switching to using Reducto's Deep Extract.

The issue isn't that models are bad at reading documents. It's that single-pass extraction has no mechanism to catch its own mistakes, and models get lazy.

For the documents that matter most, it gets to 99–100% field accuracy, even out-performing expert human labelers on extraction tasks.

The demand for these medications has been the most ferocious thing I have witnessed in my working life, and the hardest parts of running a telehealth company, like finding doctors and fulfilling prescriptions, can be entirely outsourced to platforms like CareValidate and OpenLoop.

His affiliates, armed with AI, built fake doctor profiles in Meta ads and made unscrupulous claims about weight loss using fake testimonials.

The cost of understanding what happens in a video has dropped by a factor of roughly 40, while the quality of that understanding has improved dramatically.

Historically, AI evaluation has leaned toward the forest approach. Most researchers settle for 1 to 5 raters per item, assuming this is enough to find a single 'correct' truth.

Reconstructing raw inputs forces models to model irrelevant low-level detail. Predicting in a learned embedding space allows the model to focus on semantically meaningful, causally relevant features.

Whether or not this specific bet pays off, the underlying argument that the next meaningful leap in AI capability requires moving beyond language modeling is increasingly hard to dismiss.

AMI Labs is not building a product for immediate deployment. This is a fundamental research effort, likely measured in years before commercial applications emerge.

LLMs have no grounded understanding of the physical world. They model the statistical distribution of language about reality, not reality itself.

You have to have people that have the ability to rethink the workflow at a scale that AI can execute, versus at a scale that humans can execute.

95% of organizations are getting zero return on AI deployed, with most failures found due to 'brittle workflows.'

in 2024, 47% of AI solutions were built internally and 53% were purchased; today, 76% of all AI is purchased rather than developed in-house.

Consequently, they cannot verify if tools were actually invoked, applied correctly, or used efficiently.

Experimental results show the best model, Gemini3-pro, achieves 56.3% overall accuracy, which falls significantly to 23.0% on Level-3 tasks

However, existing evaluations fall short: they lack flexible tool integration, test visual and search tools separately, and evaluate primarily by final answers.

the inherent limitations of such a single-paradigm approach pose a fundamental challenge for existing models