59
GPT-5.6一小时解开50年数学猜想，700词Prompt驾驭64个子Agent
A new prompting method using 64 multi-agent systems reportedly enabled an advanced language model to solve a 50-year-old mathematical conjecture within an hour.
Top stories in Research
42
The next frontier of AI: how ‘world models’ are simulating reality and virtual spaces
Chinese technology firms are increasingly focusing on developing world models designed to simulate physical and digital environments over time.
41
AI Scientists Are Already on the Job in Beijing: Mixing Conduction Fluid in 20 Seconds
Research laboratories in Beijing have deployed autonomous robotic systems capable of independently designing experiments and discovering new materials.
37
UltraX: Refining Pre-Training Data at Scale with Adaptive Programmatic Editing
UltraX is an adaptive programmatic editing framework designed to refine large-scale pre-training datasets efficiently to improve language model performance.
37
When the Judge Changes, So Does the Measurement: Auditing LLM-as-Judge Reliability
An empirical audit of LLM-as-judge systems shows that upgrading or changing the evaluator model leads to inconsistent scoring, highlighting reliability challenges in automated evaluation.
36
When Structured Sparse Autoencoders Learn Consistent Concepts Across Modalities
Researchers propose Structured Sparse Autoencoders to help vision-language models learn consistent, non-fragmented concepts across both visual and textual modalities.
35
Two Axes of LLM Abstention: Answer Correctness and Question Answerability
The study reveals that large language models treat answer correctness and question answerability as two distinct dimensions, suggesting that a single confidence score is insufficient for proper model abstention.
35
PLURAL: A Global Dataset for Value Alignment
Researchers have released PLURAL, a large-scale dataset based on global survey data designed to help align language models with diverse international value systems.
34
SMetric: Rethink LLM Scheduling for Serving Agents with Balanced Session-centric Scheduling
SMetric is a scheduling framework designed specifically for agentic workloads, optimizing serving efficiency by prioritizing complete agent responses and leveraging high KV-cache reuse.
34
WebSwarm: Recursive Multi-Agent Orchestration for Deep-and-Wide Web Search
WebSwarm is a recursive multi-agent framework designed to improve the depth and coverage of web search tasks by dynamically orchestrating specialized sub-agents.
33
Cognitive-structured Multimodal Agent for Multimodal Understanding, Generation, and Editing
To prevent context window overload in long-horizon multimodal dialogues, this paper introduces an agent architecture that stores visual information in an external memory structure.
33
Rethinking LLM-as-a-Judge: Representation-as-a-Judge with Small Language Models via Semantic Capacity Asymmetry
Researchers propose a Representation-as-a-Judge method that leverages the internal representations of small language models to evaluate outputs efficiently without relying on text generation.
33
Prismata: Confining Cross-Site Prompt Injection in Web Agents
The Prismata framework is designed to secure autonomous web agents against cross-site prompt injection attacks by isolating untrusted web content from core agent instructions.
33
When Implausible Tokens Get Reinforced: Tail-Aware Credit Calibration for LLM Reinforcement Learning
The paper identifies a credit assignment failure in critic-free reinforcement learning for language models and proposes a tail-aware calibration method to address it.
32
Distributed Attacks in Persistent-State AI Control
This research explores security vulnerabilities in autonomous AI coding agents, demonstrating how persistent-state environments allow agents to distribute malicious payloads across multiple pull requests.
32
A Practical Investigation of Training-free Relaxed Speculative Decoding
This study evaluates the practical trade-offs of training-free relaxed speculative decoding, analyzing how relaxing strict distribution preservation affects LLM generation speed and quality.
32
When Synthetic Speech Is All You Have: Better Call GRPO
This paper demonstrates that using Group Relative Policy Optimization helps bridge the acoustic gap when training automatic speech recognition models on synthetic text-to-speech data.
32
Mechanistic Interpretability of LLM Jailbreaks via Internal Attribution Graphs
A new method using internal attribution graphs has been proposed to analyze how adversarial prompts alter the internal reasoning of large language models.
32
Does Mixture-of-Experts Actually Help Inference on Consumer and Edge Hardware? An Empirical Study
An empirical study investigates whether the theoretical computational advantages of Mixture-of-Experts models translate to faster and cheaper inference on consumer and edge hardware.
31
Compete Then Collaborate: Frontier AI Teachers Build a Verifiable Curriculum to Improve a Coding Student Beyond Imitation
A new compete-then-collaborate framework ranks frontier AI models to build a verifiable curriculum for training smaller student models in coding tasks.
31
From Solvers to Research: Large Language Model-Driven Formal Mathematics at the Research Frontier
This paper explores the transition of large language model-driven theorem provers from solving well-defined problems to assisting with frontier mathematical research and open conjectures.
31
When LLMs Agree, Are They Right? Auditing Self-Consistency and Cross-Model Agreement as Confidence Signals
A new study demonstrates that consistency and agreement among language models are unreliable indicators of accuracy when evaluating AI outputs.
31
Peer-Predictive Self-Training for Language Model Reasoning
Peer-Predictive Self-Training is a collaborative, label-free fine-tuning framework where multiple language models use aggregated responses to mutually improve their reasoning capabilities.
31
ReCoLoRA: Spectrum-Aware Recursive Consolidation for Continual LLM Fine-Tuning
Researchers have introduced ReCoLoRA, a framework that uses spectrum-aware recursive consolidation to enable continual fine-tuning of large language models without losing performance on prior tasks.
Page 1Older →