AI Research Daily

Top AI Papers of the Week

• 🔥 Top AI Papers List: Includes AIRA, MetaCogAgent, Memory as a Model, Code as Agent Harness and Weak-Model C… from...

This week's notable AI papers roundup highlights:

• AIRA
• MetaCogAgent
• Memory as a Model
• Code as Agent Harness
• Weak-Model C…

Speculative decoding pairs a fast draft model that proposes token sequences with a single-pass verification step by the target model, delivering up to...

scpFormer introduces a generative pre-trained approach to single-cell proteomics that resolves data fragmentation via protein sequence-based...

• PhyMotion shifts video evaluation from 2D pixels to explicit 3D physics simulation, scoring kinematics, contact, and dynamics for realistic human...

DeepSeek-V4 replaces standard attention with CSA and HCA mechanisms while adding Manifold-Constrained Hyper-Connections to compress million-token...

AI functions as a research copilot in mathematics, excelling at ideation and cross-field connections while struggling with logical rigor and...

Quantitative Role Clarity cuts role overstepping in multi-agent LLM systems from 43.4% to 0.2% via a role-clarity matrix and LoRA fine-tuning. This moves beyond vague prompts to deliver measurable consistency in long-horizon tasks.

Math Reasoning Advances

• 🔥 AlphaProof Nexus: Google DeepMind's AlphaProof Nexus solves 9 Erdős problems and proves 44 OEIS conjectures using...

arXiv now issues standardized bans for submissions with hallucinated references or leftover LLM prompts, signaling authors failed to verify...

A new benchmark built from real terminal recordings shows frontier models still lag on authentic workflows.

• Engine processed 80,870 recordings into...

Three papers showcase the shift toward adaptive, value-aligned LLMs.

• Fast-Slow Training separates fixed slow weights from dynamic fast-weight...

The 2020 Lewis et al. paper introduced end-to-end training of retriever and generator with documents as latent variables and marginalization over K...

GenEvolve models image generation as tool-orchestrated trajectories that gather evidence and compose prompt-reference programs. By distilling...

• DeepMind's AlphaProof Nexus solved 9 Erdős problems out of 353 and proved 44 OEIS conjectures out of 492.
• The system pairs LLMs with Lean for...

Dan Shipper defines AI slop as visible sameness repeated endlessly, creating an "abundance of sameness" that floods content with indistinguishable outputs. This highlights how automation amplifies repetitive material rather than unique value.

Lindsey DeWitt Prat outlines a five-step blueprint to counter the "research risk cascade" where AI errors compound across pipelines like a game of...

An AI executive order scheduled for release yesterday is now available in full via public posts, opening its policy details to broader review.

Standard quantization schemes underperform on autoregressive video diffusion models because of highly unbalanced frame-wise sensitivity and...

SCRL derives verifiable subproblems from reasoning chains and applies subproblem-level normalization, converting partial progress on hard problems...