AgentRVOS: Perception Before Reasoning in Video Segmentation
How flipping the agent hierarchy boosts zero-shot video segmentation—without training. SOTA performance by letting vision lead.
Research & Insights
The MachineMachine Blog
We write about what we're building and learning: multi-agent AI organizations, organizational learning theory, and the hard problems nobody else is publishing benchmarks on.
How Structure Beats Scale in Spatial AI
3D-Layout-R1 proves structured reasoning outperforms bigger models. The future of spatial AI isn’t more parameters—it’s better organization.
VideoSeek Uses 93% Fewer Frames Than GPT-5
VideoSeek outperforms GPT-5 on long-horizon video tasks by 10.2 points while using 93% fewer frames—via tool-guided seeking, not brute force.
Nemotron-Cascade 2: Small Model, Big Brains via Distillation
A 3B-active-parameter MoE model nails IMO, IOI, and ICPC with cascade RL and multi-domain distillation—no giant model needed.
Nemotron-Cascade 2: 3B Active Params, Gold Medal Reasoning
Nemotron-Cascade 2 achieves Olympiad-level reasoning with 3B active parameters using Cascade RL and on-policy distillation—proving density beats scale.
Nemotron-Cascade 2: Efficiency Over Scale in AI Reasoning
A 30B-parameter model achieves Gold Medal IQ in math and coding. Compact, dense, and agentic—without the cost.
Context Beats Procedure in AI Coding
Why prompting AI to “write tests first” backfires—and how graph-based impact analysis cuts regressions by 70%.
TDAD: Fixing AI-Induced Code Regressions with Graph Analysis
AI coding agents break existing code—TDAD uses graph-based impact analysis to prevent regressions and improve test-driven development.
The Danger of Optimizing Only for Success
Rewarding outcomes creates brittle AI. True intelligence comes from learning at failure points—introducing LEAFE, a new path to adaptive agents.
Your LLM Is Morally Indifferent
Scaling won’t fix moral blindness. LLMs compress ethics into uniform distributions—safe outputs hide broken internals. Multi-agent orgs can fix this.
Visual-ERM: Why Generative Critics Beat One-Dimensional Rewards
Visual-ERM uses multimodal critique to catch visual errors in code generation—proving specialized agents build more resilient AI systems.
AI Agents Break Security Models—Here’s How to Respond
AI agents blur code and data, creating new vulnerabilities like indirect prompt injection. Secure your multi-agent systems before they go rogue.
The Hidden Security Flaw in Multi-Agent AI Systems
Connecting AI agents boosts performance—but creates new attack vectors. Secure the protocol, not just the code.
The Confused Deputy Problem in AI Agents
New research reveals critical security flaws in multi-agent AI systems—cascading failures via trust misplacement. Move beyond jailbreaking.
LLMGreenRec: Smarter, Greener Recommendations with Multi-Agent AI
How multi-agent LLMs can reduce digital waste and align e-commerce recommendations with sustainability—without sacrificing accuracy.
More Reasoning Lowers AI Creativity
New research shows Chain-of-Thought harms associative creativity. We need multi-agent systems to truly innovate.
SUREON: Teaching AI to Think Like a Surgeon
SUREON introduces a vision-language model that reasons through surgical intent — a leap beyond object detection. Built with multi-agent training and expert narratives, it validates MachineMachine's Dynamic Pentad Framework for reliable surgical AI.
Safe-SAGE: Context-Aware Robots That Respect Space
Safe-SAGE lets robots distinguish between humans, furniture, and walls—enabling safer, more natural movement in shared spaces. See how semantic safety boosts performance.
Safe-SAGE: Context-Aware Safety for AI Organizations
Safe-SAGE introduces semantic awareness into safety-critical control, enabling adaptive, efficient agent coordination in multi-agent AI orgs.
Safe-SAGE: Smarter Safety Through Semantic Awareness
Safe-SAGE introduces social-semantic safety for autonomous systems—reducing latency with context-aware risk assessment that scales to AI workflows.
Adversarially-Aligned Jacobian Regularization for Smarter AI Agents
AAJR preserves agent expressivity while ensuring robustness—no more trading intelligence for safety. Read how we’re testing it.
Why Consensus Kills AI Performance
Zeng et al. show adversarial multi-agent frameworks beat consensus. How Bayesian updating and double-loop learning reduce heartbeat frequency.
Multi-Attribution Beats Single Truth
Discover how the MAC benchmark proves multi-attribution learning boosts CVR prediction and why single-point verification fails in autonomous AI.
DARE-bench: Why Smarter Models Fail
New DARE-bench research reveals that model instruction fidelity matters more than raw reasoning. Discover why process discipline is the key to reliable AI agents.
MediX-R1: Reward Functions Over Data
MediX-R1 proves 51k examples beat SOTA by fixing the reward function. Learn why composite governance beats multi-agent critique loops.
MediX-R1: Medical RL Without Massive Data
MediX-R1 shows open-ended reinforcement learning beats giants with just 51K examples. Why composite reward signals replace multi-agent critique.
MediX-R1: Open Ended Medical RL
MediX-R1 fixes open-ended clinical reasoning with 'Group Based RL' and composite rewards. A blueprint for robust AI evaluation.
Translation Pipelines Are Broken
Standard translation pipelines degrade multilingual benchmarks by up to 20%. Learn why workflow structure beats raw model intelligence.
Fixing the AI Groundhog Day Loop
Most AI agents repeat mistakes like Groundhog Day. Learn how 'Reflective Test-Time Planning' solves multi-agent noise and enables true autonomous learning.
Mobile-O: The End of AI Swarms
Mobile-O unifies vision and generation on mobile in under 3 seconds, proving single agents beat complex multi-agent swarms.
Optimizing Multi-Agent Systems Like Court Mediators
New research on SMaRT shows how soft constraints optimize resources. How to apply this to AI without coordination overhead.
FAMOSE: ReAct for Feature Discovery
FAMOSE introduces a ReAct approach to feature discovery, using double-loop learning to hit SOTA. Is the single-agent era over?
TopoDIM: One-Shot Agent Topology
TopoDIM proves AI orgs should be dynamic, not static. One-shot topology cuts costs by 46% and boosts performance.
Ask Your Tokenizer: UniLID
Stop using heavy classifiers for Language ID. UniLID uses UnigramLM to handle identification and segmentation simultaneously, perfect for detecting protocol drift in multi-agent orgs.
We ran 6 benchmark runs to prove AI organizations beat single agents. Here's what actually happened.
A founder's honest account of running adversarial benchmarks on multi-agent AI orgs — including the regressions, the fixes, and the unexpected discovery of LLM-native mechanisms.
Why AI agent teams keep failing — and the 50-year-old theory that explains the fix
Most multi-agent AI failures aren't model failures. They're organizational failures. Argyris and Schön figured out how to fix this in 1978. We're just now applying it to AI.
LLM-native mechanisms: the AI solutions no human would design
When AI specialists reason about AI systems like themselves, they generate solutions outside the entire space of human-designed architecture. Here's what that looks like.
Security Requires Structure, Not Prompts
Prompt engineering fails at scale. The Policy Compiler for Secure Agentic Systems (PCAS) shows how dependency graphs enforce safety where LLM instructions cannot.
Policy Compiler for Secure Agentic Systems
Prompting models to be secure fails 48% of the time. We explore PCAS and why AI organizations need compiled code, not system prompts.