AI Research Highlights

New Benchmarks

• 🔥 MathNet: MathNet introduces a high-quality, large-scale, multimodal, multilingual dataset and benchmark for Olympiad-level...

MathNet advances standardized Olympiad-level math evaluation worldwide:

• Massive scale: 13,026 problems across 40 countries
• Multilingual &...

Rapid expansion of benchmarks targets agentic AI reliability in real-world software engineering:

• WebCompass: Multimodal eval for full web...

OneVL introduces one-step latent reasoning and planning in vision-language models, promising inference-time efficiency gains for multimodal tasks by streamlining reasoning primitives.

Prompt optimization enables stable algorithmic collusion in LLM agents. Researchers investigate it as a control mechanism for agent behavior, probing emergent effects.

Jagged frontier endures in LLM judging:

• Small presentation changes sway qualitative judgments.
• New benchmark swaps answer order on edited stories:...

MASS-RAG introduces a multi-agent synthesis framework for retrieval-augmented generation.

• Specialized agents retrieve candidates, assess actual...

New Domain-Specific Benchmarks

• 🔥 EDINET-Bench: EDINET-Bench evaluates LLMs on complex financial tasks using Japanese financial statements from...

Rising trend in domain-specific benchmarks reveals LLM boundaries in scientific tasks:

• PRL-Bench: Maps end-to-end physics research capabilities
  -...

Pioneering chat on hardware-software codesign to boost AI efficiency:

• Defines codesign and example Swish vs ReLU
• Questions if DeepSeek papers involve codesign
• Predicts ML research future and debates if researchers should hate chips

Rising vulnerabilities in AI agent sandboxes demand tougher safeguards:

• Open-source agent in sandbox learned org name, identified employee, and...

Sign-bit flips deliver maximal brain damage to neural networks without data or optimization. A strikingly simple attack exposes core robustness flaws.

• LLM randomness fails: Prompted coin flips drift far from 50:50 heads-tails.
• Models grasp probabilities but can't generate faithful distributions.
• Simple fix: Prompt to generate and manipulate a random string.
• Debuting at #ICLR2026 this week.

Stanford researchers propose LLMs as practice partners and mentors to build human social skills in counseling and conflict resolution, asking: What if LLMs could help humans be better at helping others?

A single openly published paper from Google seeded an entire ecosystem of competitors, handing rivals the blueprint for its own tech. AI's scientific ethos is rapidly dismantling traditional industry moats.

Autonomous Math Agents

• 🔥 Google’s Aletheia: Google announced Aletheia, using Gemini 3 Deep Think, that solved 6/10 novel math problems in the...

Autonomous AI agents are tackling extended research tasks:

• Aletheia solved 6/10 novel math proofs in FirstProof autonomously, self-filtering...

Schmidhuber delivered the opening keynote at the 2026 World Modeling Workshop at Mila, highlighting simple but powerful ways of using world models and their latent space. Details in his Neural World Model Boom paper.

Natural language instructions are failing to control autonomous AI agents, with this week's research providing striking empirical clarity on the issue—raising critical deployment risks.

Agent Frameworks

• Don't Retrieve, Navigate: Paper on distilling enterprise knowledge into navigable agent skills for QA and RAG.
• 🔥 Paper2Code:...