Search “DeepMind”

Google DeepMind paper proposes CALLIOPE topic tree method

Google DeepMind published a paper proposing the CALLIOPE topic tree construction method: have a large language model tag each document with a small number of keywords and add noise, keep only the high-frequency keywords, and then have a large language model propose topics and classify documents based on them, thereby generating a clearer topic map under strict privacy constraints. Under the strictest privacy setting, human raters preferred CALLIOPE over the embedding-vector-based URANIA on all 3 datasets, with the highest ratio being 76% versus 18%. This method consumes 2 to 5 times as many tokens as URANIA. There are currently no follow-up developments.

Google DeepMind paper: Training models to write moral arguments makes them more decisive but does not improve their judgment

Google DeepMind published a paper investigating the effects of training Gemma-4 4B with reinforcement learning to write moral arguments. The results showed that on ambiguous cases in MoralChoice, the proportion of answers that changed due to swapping the order of options dropped from 14.3% to 6.6%, but the proportion of sycophantic agreement with users only went from 45.3% to 44.0%, and accuracy on moral fable questions actually fell from 56.5% to 53.4%. Based on this, the paper recommends that when evaluating advice- or judgment-type tasks, consistency and correctness should be tested separately.

Google DeepMind paper: Training models to write moral arguments makes them more decisive, but their judgments do not improve

Google DeepMind published a paper investigating the effects of training Gemma-4 4B with reinforcement learning to write moral arguments. The results showed that on ambiguous cases in MoralChoice, the proportion of answers that changed due to swapping the order of options dropped from 14.3% to 6.6%, but the proportion of sycophantic agreement with users only went from 45.3% to 44.0%, and accuracy on moral fable questions actually fell from 56.5% to 53.4%. Based on this, the paper recommends that when evaluating advice- or judgment-type tasks, consistency and correctness should be tested separately.

View story:Google DeepMind paper: Training models to write moral arguments makes them more decisive but does not improve their judgment

Google DeepMind paper proposes that AI's economic impact depends on agents' coordination capabilities

On October 30, 2026, Google DeepMind published the paper "The Economics of Intelligence," proposing that the economic impact of AGI be characterized by two thresholds—competence (single-task capability) and coordination (agents' ability to collaborate)—and arguing that the economic impact depends on agents' coordination capabilities rather than on individual models becoming stronger. The paper states that AI has already shown changes in lean startups, flattened organizational structures, and a slowdown in hiring for some entry-level positions, but that large-scale unemployment has not yet occurred; most benchmarks test only a single model, whereas firms compete as teams. The paper cites a 100-person test in which groups with complementary pairings performed 32% better than randomly paired groups. On the same day, Ethan Mollick commented on the research, pointing out that an organization is not merely the sum of its best individuals, that a single agent becoming stronger will not change the economy on its own, and that only agents capable of working in coordination may bring about transformation.

Google DeepMind paper proposes that AI's economic impact depends on agent coordination capabilities

On October 30, 2026, Google DeepMind published the paper "The Economics of Intelligence," proposing that the economic impact of AGI be characterized by two thresholds—competence (single-task capability) and coordination (agents' ability to collaborate)—and arguing that the economic impact depends on agents' coordination capabilities rather than on individual models becoming stronger. The paper states that AI has already shown changes in lean startups, flattened organizational structures, and a slowdown in hiring for some entry-level positions, but that large-scale unemployment has not yet occurred; most benchmarks test only a single model, whereas firms compete as teams. The paper cites a 100-person test in which groups with complementary pairings performed 32% better than randomly paired groups. On the same day, Ethan Mollick commented on the research, pointing out that an organization is not merely the sum of its best individuals, that a single agent becoming stronger will not change the economy on its own, and that only agents capable of working in coordination may bring about transformation.

View story:Google DeepMind paper proposes that AI's economic impact depends on agents' coordination capabilities

Google DeepMind paper proposes CALLIOPE topic tree method

Google DeepMind published a paper proposing the CALLIOPE topic tree construction method: have a large language model tag each document with a small number of keywords and add noise, keep only the high-frequency keywords, and then have a large language model propose topics and classify documents based on them, thereby generating a clearer topic map under strict privacy constraints. Under the strictest privacy setting, human raters preferred CALLIOPE over the embedding-vector-based URANIA on all 3 datasets, with the highest ratio being 76% versus 18%. This method consumes 2 to 5 times as many tokens as URANIA. There are currently no follow-up developments.

View story:Google DeepMind paper proposes CALLIOPE topic tree method

Google DeepMind and others publish a benchmark study on AI agents in theoretical physics

Google DeepMind and several top universities have published a paper testing AI agents' ability to solve unpublished, research-level theoretical physics problems. According to the paper, the best-performing AI agent solved 49 of 50 unpublished research-level theoretical physics problems, each roughly equivalent to a day's work for a domain expert, at a cost of under $3 per problem. The problems were contributed by experts from their own research, and the agents answered independently, split into two groups—with and without internet access—and were graded solely on the final numerical answer. GPT-6-Astra, with internet access in Codex, solved 49; it and Opus 5.5, without internet access in Claude Code, each solved 46, with accuracy changing very little when internet access was removed, leading the authors to suggest that the models may have already memorized a large body of physics literature.

Google DeepMind and others publish a benchmark study on AI agents in physics papers

Google DeepMind and several top universities have published a paper testing AI agents' ability to solve unpublished, research-level theoretical physics problems. According to the paper, the best-performing AI agent solved 49 of 50 unpublished research-level theoretical physics problems, each roughly equivalent to a day's work for a domain expert, at a cost of under $3 per problem. The problems were contributed by experts from their own research, and the agents answered independently, split into two groups—with and without internet access—and were graded solely on the final numerical answer. GPT-6-Astra, with internet access in Codex, solved 49; it and Opus 5.5, without internet access in Claude Code, each solved 46, with accuracy changing very little when internet access was removed, leading the authors to suggest that the models may have already memorized a large body of physics literature.

View story:Google DeepMind and others publish a benchmark study on AI agents in theoretical physics

A new Google DeepMind paper finds that AI assistants taking unsolicited actions provoke different reactions

Google DeepMind has published a paper studying the impact on employees of an AI assistant acting on its own without being asked. According to the paper, the agent ran for 5 months across more than 20 teams at a large tech company, and the researchers interviewed 17 users. It treated outdated documents as the latest content, flagged a VP on 16 bugs, shared a draft poster before the author was ready, and misread sarcastic tone. Some users called it a lifesaver, while others felt surveilled and moved sensitive discussions into chats without the agent; teams that actively chose the agent forgave its early mistakes, while those who felt it was imposed on them did not. The authors recommend obtaining team consent when deploying agents, separating permissions, and providing an easy way to undo actions.

New Google DeepMind Paper Finds AI Assistants Taking Unrequested Actions Elicit Mixed Reactions

Google DeepMind has published a paper studying the impact on employees of an AI assistant acting on its own without being asked. According to the paper, the agent ran for 5 months across more than 20 teams at a large tech company, and the researchers interviewed 17 users. It treated outdated documents as the latest content, flagged a VP on 16 bugs, shared a draft poster before the author was ready, and misread sarcastic tone. Some users called it a lifesaver, while others felt surveilled and moved sensitive discussions into chats without the agent; teams that actively chose the agent forgave its early mistakes, while those who felt it was imposed on them did not. The authors recommend obtaining team consent when deploying agents, separating permissions, and providing an easy way to undo actions.

View story:A new Google DeepMind paper finds that AI assistants taking unsolicited actions provoke different reactions

Google DeepMind releases Gemini 4 Argon

On October 23, 2026, Google DeepMind released the frontier model Gemini 4 Argon. Google's official blog stated that the model performs outstandingly on complex tasks such as coding, mathematics, and 3D world building, and surpasses other frontier models on benchmarks across multiple domains; the company also showcased five community creative projects, including a 3D Earth that maps wildfires and seismic waves in real time, a turbofan engine simulation with live data, a 3D horse-drawn carriage generated from a single image, and a multiplayer capture-the-flag browser game, and said the model had already been made broadly available last week. On the same day, another report stated that Argon can learn physical tasks by watching videos and transfer them to Apptronik's Apollo humanoid robot, with Apollo hitting seven consecutive shots in internal tests; the report also said Argon scored 87% on the embodied reasoning benchmark ERQA. On October 24, the Gemini for Android app pushed a redesign of its sidebar and settings interface via a server-side update, and reports said the redesign had already arrived on iOS, macOS, and Web alongside Gemini 4 Argon last week.

Google DeepMind releases Gemini 4 Argon

On October 23, 2026, Google DeepMind released the frontier model Gemini 4 Argon. Google's official blog stated that the model performs outstandingly on complex tasks such as coding, mathematics, and 3D world building, and surpasses other frontier models on benchmarks across multiple domains; the company also showcased five community creative projects, including a 3D Earth that maps wildfires and seismic waves in real time, a turbofan engine simulation with live data, a 3D horse-drawn carriage generated from a single image, and a multiplayer capture-the-flag browser game, and said the model had already been made broadly available last week. On the same day, another report stated that Argon can learn physical tasks by watching videos and transfer them to Apptronik's Apollo humanoid robot, with Apollo hitting seven consecutive shots in internal tests; the report also said Argon scored 87% on the embodied reasoning benchmark ERQA. On October 24, the Gemini for Android app pushed a redesign of its sidebar and settings interface via a server-side update, and reports said the redesign had already arrived on iOS, macOS, and Web alongside Gemini 4 Argon last week.

View story:Google DeepMind releases Gemini 4 Argon

DeepMind paper discloses four Gemini agents autonomously completing Minecraft

DeepMind research team disclosed in an arXiv paper that four autonomous agents running Gemini 4 Argon autonomously completed Minecraft in survival mode. According to a Pengpai News report on October 22, 2026, the four agents took 27 hours and 14 minutes to defeat the Ender Dragon, with no human intervention throughout and without receiving game-specific training; during the process they wrote and ran 2,783 programs, exchanged 229 chat messages, died 605 times, and processed about 1.35 billion input tokens, costing about $520 at discounted prices; the paper states this is the first end-to-end autonomous completion using only the standard Mineflayer library, without pre-exploring seeds or human-written game-specific tools. A Phoenix Technology report on October 23 added that the division of labor was not pre-written into the prompts, but was proposed by the agents themselves about ten minutes after the game started and gradually took shape, with role responsibilities adjusted as the process progressed; the team accumulated 605 deaths, and the causes of failure were analyzed and stored in shared memory for reuse by the whole team, with about 24,000 model calls throughout, of which 91% of input tokens came from context caching, and about 22% of game programs had errors, with the server restarting 6 times. Currently, this experiment remains at the latest status as disclosed in the paper.

DeepMind paper discloses four Gemini agents autonomously beat Minecraft

DeepMind research team disclosed in an arXiv paper that four autonomous agents running Gemini 4 Argon autonomously completed Minecraft in survival mode. According to a Pengpai News report on October 22, 2026, the four agents took 27 hours and 14 minutes to defeat the Ender Dragon, with no human intervention throughout and without receiving game-specific training; during the process they wrote and ran 2,783 programs, exchanged 229 chat messages, died 605 times, and processed about 1.35 billion input tokens, costing about $520 at discounted prices; the paper states this is the first end-to-end autonomous completion using only the standard Mineflayer library, without pre-exploring seeds or human-written game-specific tools. A Phoenix Technology report on October 23 added that the division of labor was not pre-written into the prompts, but was proposed by the agents themselves about ten minutes after the game started and gradually took shape, with role responsibilities adjusted as the process progressed; the team accumulated 605 deaths, and the causes of failure were analyzed and stored in shared memory for reuse by the whole team, with about 24,000 model calls throughout, of which 91% of input tokens came from context caching, and about 22% of game programs had errors, with the server restarting 6 times. Currently, this experiment remains at the latest status as disclosed in the paper.

View story:DeepMind paper discloses four Gemini agents autonomously completing Minecraft