What could become important next

Google researchers propose a method against test memorization in AI agents

Research from Google Cloud AI Research addresses the problem where AI agents that optimize their own harness overfit too much to tests. The new RRSI method regulates optimization with a budget of edit changes and a filter against task-specific tricks. Result: generalization is preserved without increasing compute costs.

NASA and IBM released an open-source foundation model for lunar science

NASA and IBM released the Lunar Foundation Model, an open-source foundation model for lunar science. Trained on 2 million tile bundles from 17 years of data from the Lunar Reconnaissance Orbiter and other missions. The model specializes in ice and crater detection.

Part of an overview of multiple AI topics; only this event has been covered.

Chinese AI models are subject to state doctrine, study by Aleph Alpha finds

Aleph Alpha created a benchmark covering 967 politically sensitive topics. Tests of the Qwen, DeepSeek, and Kimi models showed that only 17–41% of responses were balanced; the rest repeat state doctrine, evade the question, or refuse to answer. The bias also shows up in unrelated questions. Nvidia Nemotron contains approx…

ThinkingBox benchmark exposes AI agent errors by checking the final database state

The ThinkingBox benchmark evaluates AI agents across 507 business processes based on the final state of the systems. According to the authors, 79 853 out of 121 680 attempts with 12 models failed; 67.24% of the failed attempts nevertheless ended without a final tool error and still performed a state change.

Researchers propose Artificial Symbiotic Intelligence as an alternative to the singularity

In an essay, Benjamin Bratton, Blaise Agüera y Arcas and James Manyika propose Artificial Symbiotic Intelligence: an interconnected system of humans and AI agents. According to the authors, AGI may emerge through their collaboration; this is a conceptual argument, not an announcement of a new model.

LEGO-Anything project turns photographs into 3D scenes, but agents make geometry errors

Researchers from University of Maryland and AWS introduced LEGO-Anything: an agent progressively creates and edits Blender code based on a photograph. In LEGO-Bench with 208 images, the GPT-6 Astra model achieved accuracy of 53.4 % indoors and 39.6 % outdoors, according to the study.

According to the report, Anthropic is consulting religious experts on the possible consciousness of Claude models

According to a New York Times report, Anthropic has since autumn 2025 invited dozens of theologians and philosophers to discussions about the possible consciousness and moral behavior of Claude models. The report is based on interviews with 20 participants; this does not prove consciousness of the models.

Researcher Cathy Wu explores the use of reinforcement learning in transportation systems

A profile of MIT researcher Cathy Wu describes her focus on machine learning and reinforcement learning for transportation systems. According to the researcher, these methods could make it easier to compare many design variants; the available text does not provide specific results.

Study of 45 workers describes the benefit of AI in challenging one's own ideas

The study's authors interviewed 45 knowledge workers at a US public university and analyzed their conversations with AI. According to the authors, deliberately seeking out conflicting viewpoints using AI can improve problem-solving; the outputs must be assessed using one's own expertise.

Mercor study: AI handles structured accounting tasks, closing requires supervision

According to a study by Mercor, AI models outperform 12 licensed accountants on simplified tasks. In the full test with 160 tasks, the Claude Opus 5.5 model leads with 61.8% of criteria met. Nearly 60% of tasks were not fully solved by any model; closing still requires supervision.

The author of the commentary questions the reasoning ability of current LLMs

The author of the commentary explains that the AlphaGo system combined neural networks with search through possible moves. He claims that current LLMs lack similar reasoning abilities and that these are needed for reliable results in science and medicine. The text is incomplete.

The ServiceNow CoreAI team introduces AutoSynthData for training enterprise agents

The ServiceNow CoreAI team describes AutoSynthData, which generates and verifies training tasks for enterprise agents based on the failures of the target model and the successes of a stronger model. The available excerpt does not provide measurement results or a tool release.

Study presents Mem++ for preserving document history in LLM agent memory

The study presents Mem++, which stores entire documents with date and author without a generative model at write time. When queried, it selects time-matching documents. According to the authors, it outperformed the strongest comparison memory system on OrgMemBench by 8.0 to 13.1 points.

SELF-POT study measures the benefit and cost of test-time computation

The SELF-POT study compares five low-cost models across 350 tasks and counts all calls toward the cost. In repeated evaluation of coding candidates, selection based on public examples increased the number of correct solutions from 376 to 453 out of 500 and reduced API costs by 12–49%.

Qwen3 models failed to meet reliability thresholds for four AI agent subtasks

The study states that none of the 16 Qwen3 model configurations met the predetermined thresholds for four AI agent subtasks. Even 4-bit quantization failed to produce a satisfactory configuration. The authors recommend combining small models with a proven baseline procedure.

DSB-DG benchmark reveals loss of document fidelity in voice agents

The study presents the DSB-DG benchmark with 1 636 question-answer pairs from 50 documents across five professional domains. According to the authors, the accuracy of voice agents decreases with context and dialogue length; the ASR-LLM-TTS architecture performs best.

Study introduces CTWM for more efficient memory handling in language agents

The study proposes CTWM, which allocates the context budget according to memory usage frequency. The authors report a token savings of 24.48% on LongMemEval with comparable overall accuracy, and on the Synthetic Graph World a 13.6% decrease in prediction error for the less frequently used half of memory.

Study links the Qwen3.6-35B-A3B model with geometric tools for spatial planning

The authors tested an agent using the Qwen3.6-35B-A3B model with a library of geometric trajectories in a dynamic 2D environment. According to the study, the configuration without reasoning achieved the same goal-reaching success rate as the chain-of-thought variant, while cutting decision time from minutes to seconds.