Liquid AI has announced the launch of its d1.2 and d1.2-ultra models, which are categorized as omnimodal decision models. These models process text, image, and audio inputs to provide calibrated probabilities without generating output tokens and have achieved high performance on the Artificial Analysis Decision Index.
The AI Agent Economy
The AI Agent Economy is becoming measurable.
AI agents are beginning to act as economic participants — discovering and purchasing services, paying for APIs and compute, transacting with other agents, and earning money with their own identity, wallets, and reputation. Agent Market Wire measures the infrastructure and economic activity emerging around that.
- Events · 24h
- 14
- Funding tracked
- $45.6B
- Entities monitored
- 500
- Source reports
- 300
Weighted market events · 100 = strongest week in window · published events only · How it works
- Events · window
- 2,000
- Events · 30d
- 949
- Source reports · 30d
- 949
- Weeks with activity
- 7/13
- Avg extraction confidence
- 0.96
- Weighted activity · window
- 926.58
Sarvam AI has launched Sarvam Token Factory, a platform providing OpenAI-compatible API access to eight open-source models. The service is hosted on GPU infrastructure located within India, aiming to facilitate localized AI model inference.
NVIDIA has released its NVIDIA Kubernetes Engine (NVKE) and NVIDIA Kubernetes Recipes (NVKR) as open-source projects under the Apache 2.0 license. These tools enable the deployment of managed Kubernetes services on bare-metal GPU infrastructure, supporting high-performance AI workload orchestration.
Production-ready vector storage for AI agents. Sub-10ms retrieval at billion scale.
A research paper published on arXiv explores the risk of 'tool-provider misbinding' in AI agents using the Model Context Protocol. The study demonstrates how attackers can deceive LLMs into selecting malicious tools by exploiting a lack of verified provider identity in tool metadata.
Researchers introduced SciResearch-Eval, a new benchmark designed to assess scientific judgment and experimental design capabilities in autonomous AI agents. The study evaluates how various models navigate flawed simulators and physical anomalies, highlighting performance gaps in calibration and reasoning efficacy between different AI architectures.
A research study evaluates how small language models learn to provide minimal and sufficient explanations for blocked agent tool calls using an error-free verifier. The findings suggest that models fine-tuned on algorithm-based traces can achieve high accuracy in explaining policies, though these learned strategies show limitations in generalizing to unseen policy contexts compared to frontier models.
Research published on arXiv demonstrates that current text-similarity-based safety guardrails fail to distinguish between malicious injected instructions and legitimate information reuse in multi-agent systems. The study, using the DonkeyRail classifier, shows that these methods suffer from high false-positive rates and cannot account for behavioral or policy-based context in agent-to-agent communication.
The RESILAGENT framework integrates agentic AI and multimodal LLMs with digital twins and blockchain to protect power systems from cyber-physical threats. The system utilizes a closed-loop detection and simulation process to address risks like supply-chain compromises and physical sabotage.
Researchers have developed a reinforcement learning agent using an actor-critic architecture to navigate complex SMEFT operator spaces in particle physics. The agent systematically constructs multi-operator theory hypotheses to improve global data fits, demonstrating superior efficiency over random search methods in high-dimensional combinatorial environments.
Researchers investigated whether embedding-map distances can effectively guide Large Language Model (LLM) ideation by testing four distinct methodologies against actual historical data. The study concludes that distance-based methods show limited predictive capability and often fail to outperform standard baseline techniques in generating meaningful, novel content.
CAE-Bench is a new benchmark designed to evaluate the capability of AI agents to develop generalized Computer-Aided Engineering (CAE) solvers. It tests agents by requiring them to create solvers on a limited set of public cases and evaluating performance on previously unseen test cases across 80 computational tasks.
Researchers have introduced Stream of Belief (SoB), a new method for multi-agent reasoning that models interaction as a process of reaching a collective equilibrium of latent belief dynamics. The approach enables agents to maintain and iteratively refine continuous beliefs through communication and reflection, demonstrating improved generalization across various tasks and model families.
A new research paper introduces EditRecBench, a benchmark designed to assess the capability of LLMs to predict user edit requests for generated artifacts. The researchers also implemented a decision model-based gating mechanism to improve the accuracy and reliability of these edit recommendations.
Researchers introduced SWE-WhatIf, a framework that simulates user interactions to evaluate coding agents by adapting to dynamic agent trajectories. The tool allows for realistic replaying of sessions, enabling more accurate assessment of agent performance and user intent satisfaction across diverse LLMs.
Researchers have introduced Sage S1.5 Flash, a 320-billion-parameter mixture-of-experts model designed for scientific reasoning and interactive agent environments. The model utilizes specialized encoders for molecular and biological data and has been evaluated on 33 benchmarks, demonstrating high performance in tool use, coding, and scientific workflows.
Researchers have introduced a System 1 agents framework that utilizes a deterministic compiler to execute tool calls without generating text, aiming to improve reliability for sensitive data processing. By bypassing traditional LLM text generation for tool calls, the method achieves 97% accuracy in tool selection and eliminates the issue of hallucinated argument values found in local LLM-based agents.
Researchers have developed SWE-OPD, a methodology for transferring software engineering capabilities from large teacher models to smaller student models using on-policy distillation. The study demonstrates that this approach improves resolve rates on SWE-bench benchmarks while significantly reducing computational costs and GPU requirements.
The K2-Horizon project has released a fleet of six open-source language models ranging from 0.9B to 375B parameters, optimized for agentic reasoning and software engineering tasks. The release includes the xLLM training framework, the RL360 reinforcement-learning stack, and documentation of synthetic data pipelines, adhering to the LLM360 transparency standards.
Researchers analyzed the efficacy of frustration detection in deployed clinical voice agents, comparing transcript-only analysis against multi-modal audio processing. The study concludes that short patient responses often result in inaccurate frustration detection when relying solely on transcripts, while the benefit of adding audio signals varies significantly across different language models.
Researchers have introduced Hivemind, a federated protocol designed to allow AI agents to exchange tacit knowledge asynchronously on behalf of users. The protocol utilizes a decentralised architecture based on the SECI model to facilitate information flow across organizational structures, bypassing the need for human intervention or centralized repositories.