AI Frontier Daily

Google DeepMind's AlphaProof Nexus system solved dozens of long-standing mathematical conjectures, including 9 Erdős problems, and passed peer review in Science. This marks a substantive breakthrough in autonomous AI mathematical reasoning.

Built on Gemini 3.1 Pro, the system uses the Lean formal proof language for real-time error feedback iteration. It can independently solve high-difficulty problems without relying on the specialized AlphaProof tool.

AlphaProof Nexus resolved two Erdős problems pending for 56 years and 44 OEIS integer sequence conjectures. Its architecture generates drafts and corrects proofs using a formal checker, ensuring rigorous validation.

Unlike previous AI math works that were preprints or controversial, this study underwent strict peer review, establishing new standards for AI-assisted mathematical discovery.

PhAI Labs launched RSI-Anything, a framework introducing world models to predict action consequences and iterate via real feedback, breaking traditional Recursive Self-Improvement (RSI) limits. It significantly boosts success rates in code …

Experiments show that in code repair tasks, using world models to assist solution selection raised success rates from 26.3% to 31.2% without changing the base model, proving the efficacy of external optimization mechanisms.

A small 4B parameter model, after five generations of self-improvement, outperformed reference systems using larger model combinations in lab planning tasks. This indicates efficient evolution strategies can compensate for parameter gaps.

The core of RSI-Anything lies in predicting consequences before action and optimizing predictors through real feedback, enabling cross-domain self-evolution capabilities.

Researchers proposed Direct Hidden State Alignment (DHSA), using Residual Competition Maps (RCM) to quantify internal computational support or opposition to preferences. This enables dynamic inference-time intervention without modifying bas…

DHSA reveals the gap between 'output matching preferences' and 'internal consistency'. Studies found that even when outputs align with preferences, internal computations may still oppose targets, a competition not fully eliminated by method…

By introducing 256-16384 control parameters, DHSA achieves DPO-comparable alignment while preserving base model diversity. This offers a new paradigm for lightweight preference alignment.

The technique maps external preferences back to native inference computations, quantifying component support or opposition to goals, thereby directly intervening in hidden states during inference.

Microsoft CEO Satya Nadella announced the third-generation Maia 300 AI inference chip, planned for full deployment in 2027 to replace Nvidia and AMD hardware as Azure's core infrastructure. It aims to reduce large-scale AI inference costs a…

Built on TSMC N2P 2nm process with multi-chiplet design, Maia 300 claims 40% higher token throughput per megawatt than mainstream GPUs at equal power. It features 8 HBM4 stacks totaling 288GB capacity and 18TB/s bandwidth.

With 13.8 PFLOPS FP4 compute and 950W rated power, the architecture optimizes for long-context inference, low-latency chat, and multi-step agentic tasks. A single rack integrates 96 accelerators, achieving 7x token processing efficiency ove…

Maia 200 is already deployed in Copilot. Maia 300 focuses on solving data movement bottlenecks in LLM inference, leading GPU solutions by 40% in energy efficiency, marking deepened Microsoft silicon strategy.

Andrej Karpathy highlights a severe 'funnel-shaped' cognitive gap regarding LLM capabilities. Approximately 75% of the global population has barely touched LLMs, 20% use them merely as enhanced search tools, while only 0.2% experience front…

This tiny group of professional developers witnesses agents achieving qualitative leaps in math and code, compressing weeks-long complex engineering tasks into single-prompt end-to-end delivery. This reveals a vast gap between public percep…

Cross-verified sources indicate the core mechanism of this cognitive difference is task 'verifiability'. Domains like code and math have clear correctness standards (e.g., unit test pass rates), allowing reinforcement learning to efficientl…

In contrast, unstructured domains lack automatic verification mechanisms. Limited by free/paid tier restrictions and training resource tilts toward B2B commercial value, the general public struggles to intuitively perceive AI's disruptive e…

Amazon is in advanced talks to acquire AI infrastructure startup Decart for approximately $7 billion, signaling intensified strategic competition among tech giants for underlying compute software stacks. If completed, the deal will help AWS…

Decart's core asset is its 'Dynamic Operating System' (DOS) optimization stack, which claims seamless switching across heterogeneous hardware like Nvidia, Google TPU, and Amazon Trainium, achieving up to 8x model runtime acceleration.

The acquisition price represents a significant premium over Decart's May valuation of nearly $4 billion and exceeds Anthropic's previous $6 billion offer. Notably, AWS granted Decart early access to Trainium3 in 2025, establishing deep tech…

Anthropic previously abandoned a $6 billion acquisition due to due diligence issues, seeking other paths to resolve compute bottlenecks. Amazon's move aims to break compute lock-in and strengthen its position in real-time world models and p…

Huawei Noah's Ark Lab and UCL jointly launched Memento 3, an agent system that achieved perfect scores on all 25 games in the ARC-AGI-3 benchmark without any game-specific fine-tuning, keeping Claude Opus 5 weights completely frozen.

The system boosted Relative Human Action Efficiency (RHAE) from a baseline of 40.7 to 100, reducing total operations by approximately 56% compared to human averages, achieving extreme efficiency breakthroughs in complex interactive tasks.

Memento 3's core innovation is 'recursive self-improvement based on world models'. Instead of relying on pre-trained knowledge or parameter updates, the AI constructs checkable, executable, and rollback-capable external state graph environm…

This architecture allows agents to rapidly converge on optimal strategies in unfamiliar environments. In Atari Pong tests, it demonstrated the ability to autonomously iterate strategies without real-time LLM inference, significantly reducin…

NTU and other institutions proposed RoboFoundry, formally defining embodied agent systems as 'self-evolving System-as-Policy'. The framework enables autonomous optimization by automatically diagnosing and evolving contexts, skills, and exec…

Experiments showed GPT-5.5's success rate on EmbodiedBench rose from 56.9% to 72.7%, significantly surpassing baselines in long-term memory and perturbation tests. This proves system-level evolution can yield performance leaps without modif…

RoboFoundry defines a new paradigm: treating contexts, skills, and execution interfaces as independently evolvable components rather than static configurations. This 'System-as-Policy' concept breaks traditional reliance on model weights.

The framework demonstrated zero-shot transfer across robots and tasks, plus post-deployment online evolution on real robots. In LIBERO-PRO perturbation tests, it outperformed CaP-Agent0 by 243.8–679.7%.

Addressing the pain point where fixed-threshold compression causes loss of key intermediate results in long agent trajectories, Apple ML Research proposed the Self-Compacting framework. It introduces compression tools and lightweight Rubric…

Unlike traditional fixed-interval compression relying on external triggers, Self-Compacting allows models to judge compression timing based on task status (e.g., subtask resolution or convergence), avoiding discarding partial results mid-de…

Experiments show the method significantly reduces cost per question across six benchmarks and seven models. It reveals a metacognitive gap where models lack self-awareness of context rot, proving scaffolding can substitute for training to s…

The core mechanism is 'compression tool + lightweight Rubric'. The Rubric clearly defines when to trigger compression (subtask resolution/convergence) and when to suppress it (mid-derivation/stuck), enabling smarter context management.

vLLM implemented bidirectional KV Cache transmission in Prefill-Decode disaggregated architectures, addressing high TTFT latency caused by recomputing historical tokens in multi-turn Agent scenarios. The solution allows Decode nodes to send…

Traditional unidirectional KV transmission forces Prefill nodes to recompute tokens generated in previous turns during multi-turn dialogues, increasing TTFT latency. Bidirectional transmission enables Decode nodes to share complete session …

This optimization is particularly suitable for complex Agent workloads with frequent tool calls and accumulating context. By reusing KV states, it significantly reduces inference overhead associated with context accumulation.

This improvement enhances multi-turn interaction efficiency, providing underlying infrastructure support for building low-latency, high-throughput agent applications.