Tagged articles
benchmark evaluation
57 articles · Page 1 of 1
OpenAI launches GPT-6.2 Astra for complex professional tasks and Sol for high-frequency general work, detailing pricing, benchmark results, evaluation divergences, tool-calling risks, safety thresholds, and a practical routing matrix for production deployment.
AI SafetyAstraGPT-6.2
0 likes · 13 min read

A survey by 80 researchers from 37 institutions analyzes 299 multimodal agent memory architectures and 82 benchmarks, introducing a dual-axis framework and three-tier classification (static, dynamic, self-evolving) to advance memory system design.
AI agentsMemory ManagementMultimodal Agent Memory
0 likes · 7 min read

Researchers from NUS, UCB, and HKUST propose Route Anything, a unified routing framework where a 4B parameter model selects among diverse candidates (LLMs, agents, robot policies) per input, outperforming fixed strong models like Opus 5.5 on Terminal-Bench (77.78 vs 69.44) and cutting API costs by 96% with minimal quality loss.
LLM routingRoute Anythingbenchmark evaluation
0 likes · 11 min read

Peking University researchers release OmniEdu, an open-source K-12 educational foundation model family (4B, 9B, 27B) built on Qwen3.5/3.8, featuring curriculum alignment, diagnostic reasoning, and pedagogical scaffolding, with full paper, weights, training data, code, and evaluation toolkit, outperforming base models on education benchmarks while retaining general capabilities.
FP8 quantizationK-12 educationOmniEdu
0 likes · 14 min read

Jilin University and Shanghai AI Lab propose EvoTeam, a multi-agent framework where agent skills, team composition, and interaction rules co-evolve across tasks, enabling adaptive team formation, team-scoped collaboration, and multi-level experience reuse, outperforming baselines on GAIA, LoCoBench, and BeyondSWE with cross-model transfer potential.
Continuous EvolutionCross-Model TransferEvoTeam
0 likes · 8 min read

Researchers from InfRec, Xidian, Tongji, and NUS open-source The Trinity, a multimodal recursive self-improvement framework that shifts the evolutionary unit from tasks to capabilities, enabling agents to share, inherit, and continuously optimize improvements across diverse tasks, demonstrating 34.1 and 26.2 point gains on 28 training and 18 unseen benchmarks after four RSI rounds, with transfer to real robot tasks.
Agent HarnessCapability-Driven EvolutionMultimodal Agents
0 likes · 14 min read

CUA-Master-27B, a 27B-parameter computer-use agent from CUHK MMLab and JD Research, learns from software tutorial videos via a dual recursive framework, achieving 43.9 on WeaveBench (near GPT-5.6's 44.0) with 28% less execution time, and 25% pass rate on Agents' Last Exam with 34-36% speedups over Kimi-K3 and Qwen-3.8-Max.
Agent Training DataCUA-MasterComputer-Use Agents
0 likes · 9 min read

The author tests Tencent's Hunyuan Hy4 model across six demanding agent workflows in WorkBuddy — video generation from docs, lecture-to-webpage conversion, game cloning, 3D teardown simulators, industry report generation, and templated PPT creation — finding Hy4 handles 30-40 minute long-horizon tasks with solid code, vision, and planning at $0.55/task, 62% cheaper than Kimi K3.1 Pro.
AI agentsHunyuan Hy4WorkBuddy
0 likes · 8 min read

Ant Group open-sources AntOmniEvo, a framework that automates AI system optimization by evolving prompts, rules, and configurations through a self-improving loop with the Mara Chain algorithm and Pareto-filtered selection, achieving state-of-the-art results on AppWorld, TerminalBench, and MuSiQue benchmarks with 3x sample efficiency.
AI system optimizationAntOmniEvoMara Chain
0 likes · 17 min read

The article introduces Harness-of-Harness (HoH), a protocol organizing AI coding agents into iterative planning, development, and independent QA cycles with persistent artifact and evidence states, showing 52.25% average improvement over vanilla agents across three benchmarks and demonstrating a 70-round FPS game development case study.
AI coding agentsHarness-of-HarnessMulti-Agent Systems
0 likes · 17 min read

IQuest Research releases technical report for IQuest-Q1, a 320B MoE model with 15B active parameters, 512K context window, recursive multi-token prediction, and agent-focused post-training, achieving strong benchmarks across coding, tool use, and reasoning tasks.
Mixture of ExpertsRecursive MTPagent training
0 likes · 16 min read

Shanghai AI Lab's Harness-of-Harness (HoH) framework organizes AI coding agents into planning, development, and QA roles with persistent state across iterations, achieving 52.25% average improvement over vanilla agents on three benchmarks and demonstrating 70-round autonomous FPS game development with 81 tracked issues.
AI coding agentsHarness-of-HarnessMulti-Agent Systems
0 likes · 18 min read

Renmin University introduces AEWM, an editing world model that classifies and revises noisy agent actions in real time, improving long-horizon task performance across search, terminal, and software engineering benchmarks without fabricating tool feedback.
AEWMAgent CorrectionEditAct
0 likes · 8 min read

Evolvent AI and NUS researchers open-source RSIGym, an Everything-as-a-Service research environment for recursive self-improvement, introducing the RSI-Index metric to evaluate agents' ability to jointly optimize model weights and execution harnesses, with experiments showing Opus 5 achieving 0.4809 RSI-Index and significant gains on SWE-bench and AIME benchmarks.
AI agentsEverything as a ServiceModel-Harness Co-optimization
0 likes · 20 min read

OpenRSI Index evaluates AI agents on real open-source research projects — improving Marin's optimizer, Qwen's RL tasks, and GPIC image generation — using thousands of H100 hours, with isolated workspaces and independent verification to ensure reproducible progress beyond benchmark gaming.
AI research agentsGPU computeOpenRSI-Index
0 likes · 13 min read

EvoOntology introduces a self-evolving ontology layer for data agents that transforms static semantic layers into a runtime service, using a three-layer architecture and an evolution loop where agents' execution traces drive continuous ontology updates via builder and evolution agents with validation gates, reducing token usage and improving accuracy across benchmarks.
Data AgentsEvoOntologyLLM agents
0 likes · 18 min read

Multimodal AI can accurately extract text from table screenshots but often fails to grasp structural relationships like merged headers, units, and footnotes, leading to fluent yet misleading summaries; research shows text recognition accuracy does not equal reliable business interpretation, so AI should provide verifiable observations with source references for human verification.
AI reliabilityHuman-in-the-Loopbenchmark evaluation
0 likes · 8 min read

EverMind's Raven V0.2.0 introduces a Harness of Harnesses architecture that orchestrates specialized agents like Raven-Research, Raven-Code, Raven-Design, Raven-Oncall, Claude Code, and Codex via DAG-based task graphs, while its experimental Curator component enables runtime Harness self-modification across Memory, Planning, Capability, and Action policies, demonstrating superior benchmark results in multi-agent orchestration, deep research, coding, design, and continuous execution tasks.
AI agent frameworkDAG task graphEverMind
0 likes · 26 min read

OpenAI launched GPT-6 Sol and Luna while Anthropic released Claude Opus 5.5, all emphasizing lower costs and improved capabilities; the article compares pricing, benchmarks, cache economics, and independent evaluations, advising developers to test models on their own tasks to verify real-world cost per qualified result.
Claude Opus 5.5GPT-6LLM pricing
0 likes · 14 min read

Google introduces Procedural Graphs, a self-evolving graph structure that explicitly encodes procedural knowledge for LLM agents, enabling them to locate, extract, and generate contextual guidance from a (procedure, relation, procedure) triplet graph, achieving 85% survival from 0% in CFO simulation and winning 21 of 24 model-benchmark combinations across 7 benchmarks and 4 LLMs.
Agent ReliabilityGraph-based ReasoningLLM agents
0 likes · 9 min read

This article reveals three major sources of waste in AI agent workflows — redundant context recomputation, outdated prompt patterns, and misallocated reasoning effort — and provides a systematic optimization framework with caching strategies, prompt auditing, effort calibration, and holdout-set validation, demonstrating 50–73% cost reductions without performance loss across benchmarks.
AI EngineeringPrompt Cachingagent optimization
0 likes · 25 min read

This article analyzes MemoHarness, a framework that evolves AI agents by optimizing six external harness dimensions—context assembly, tool interaction, generation control, task orchestration, memory management, and output processing—instead of model weights, demonstrating gains on terminal, coding, and finance tasks while noting limited experimental scale and selective cross-task transfer.
Agent HarnessAgent MemoryExternal System Evolution
0 likes · 22 min read

This article presents a reproducible, three‑dimensional deterministic evaluation framework (outcome, compliance, process) built on Alibaba Cloud AgentLoop, applies it to a 10‑task subset of terminal‑bench 2.1 to benchmark DeepSeek Harness against Codex, details the methodology, results, and future research directions.
AI agent assessmentAgentLoopDeepSeek Harness
0 likes · 26 min read

The article analyzes Turing DeepVision’s new multi‑modal model TuringJian X, detailing its LLM‑EDC architecture, evidence‑driven training, diverse product‑category capabilities, and a rigorous benchmark that shows superior accuracy, speed, and explainability over traditional human platforms.
AI authenticationLLM-EDCRLVR
0 likes · 20 min read

CoLT replaces the verbose textual chain‑of‑thought with just three latent vectors, delivering up to 22.6× faster generation and 10.1× end‑to‑end inference speedups while achieving a 79.1% average accuracy across eight multimodal benchmarks, all without any auxiliary visual annotations.
Chain of Latent ThoughtsEfficiencybenchmark evaluation
0 likes · 14 min read

The article analyzes recent AI updates—from DeepSeek's DSpark inference boost and FlashAttention‑4's kernel redesign to Codex UI tweaks and design‑mode tools—arguing that the competition is shifting from answering questions to actually completing tasks, and it highlights three layers of progress, evaluation challenges, and the practical questions we must now ask of AI agents.
AIAgent AutomationDeepSeek
0 likes · 19 min read

The paper shows that many failures of deterministic LLM agents stem from mismatched model‑environment interfaces, and introduces LIFE‑HARNESS—a four‑layer runtime harness that extracts reusable failure patterns from training trajectories without updating model weights, delivering an average 88.5% relative performance gain across 126 model‑environment settings.
Deterministic AgentsLLM agentsLife-Harness
0 likes · 8 min read

The article analyzes why existing audio‑video QA pipelines break narrative continuity, proposes a structured‑script and evidence‑chain approach to automatically build the OmniVideo-100K dataset of 100K high‑quality QA pairs, and shows that fine‑tuning open‑source multimodal models on this data yields consistent accuracy gains across multiple benchmarks.
OmniVideo-100Kaudio-video datasetbenchmark evaluation
0 likes · 12 min read

The paper introduces Deep Pre‑Alignment (DPA), a novel Vision‑Language Model architecture that inserts a perceiver VLM to pre‑align visual features with the LLM’s text space, reducing alignment cost, preserving language ability, and delivering consistent multimodal performance gains across multiple benchmarks with minimal inference overhead.
Deep Pre-AlignmentLLMPerceiver VLM
0 likes · 10 min read

NEWTON introduces an Agent‑centric framework that augments existing video generators with a planner, physics‑aware tools, and a verification loop, enabling multi‑round refinement and significantly improving physical consistency on benchmarks without retraining the underlying generator.
Agentic AITool OrchestrationVideo Generation
0 likes · 8 min read

The article introduces LIFE‑HARNESS, a runtime‑interface adaptation framework that keeps model weights unchanged, extracts reusable failure patterns from a single model's training trace, and achieves an average 88.5% relative performance gain across 18 LLM agents and 7 deterministic environments, with successful transfer to 17 other models.
Cross-Model TransferLLM agentsRuntime Harness
0 likes · 8 min read

The article reviews the "Natural-Language Agent Harnesses" paper, explains the distinction between code, middleware, and harness layers for LLM agents, introduces NLAH and IHR concepts, and details experimental evaluations that show natural‑language harnesses can match code‑based control while exposing new trade‑offs and risks.
Agent OrchestrationIntelligent Harness RuntimeLLM agents
0 likes · 13 min read

The article examines the PhoneWorld project, which reconstructs realistic Android app environments from user interaction traces to create scalable, resettable, and verifiable mock apps, enabling large‑scale training and evaluation of Mobile Agents with demonstrated performance gains across multiple benchmarks.
Mobile AgentPhoneWorldapp generation
0 likes · 12 min read

SkillOpt reframes LLM agent skills as trainable external state, applying a deep‑learning‑style optimizer to systematically improve skill documents, and demonstrates across six benchmarks, seven models, and three execution modes that this approach yields consistent, large gains and robust transferability.
SkillOptText‑Space Optimizationagent skills
0 likes · 12 min read

SkillOpt treats an agent's skill file as a trainable text document, iteratively editing it with a frozen LLM to boost performance on tasks with known correct answers, achieving large benchmark gains without fine‑tuning model weights while remaining portable and auditable.
AI agentsLLM training alternativePrompt Optimization
0 likes · 18 min read

The SE‑GA framework introduces hierarchical memory and a two‑stage self‑evolution training pipeline that enable GUI agents to retain critical context, learn from past successes, and achieve state‑of‑the‑art performance on long‑horizon benchmarks.
GUI agentsbenchmark evaluationmemory-augmented AI
0 likes · 11 min read

Thought-Aligner introduces a lightweight, plug‑in safety layer that corrects unsafe reasoning in AI agents during the millisecond window between thought generation and action execution, dramatically improving behavioral safety while preserving task usefulness across benchmark and real‑world deployments.
AI Safetyagent alignmentbenchmark evaluation
0 likes · 11 min read

Microsoft’s newly open‑source SkillOpt framework treats an agent’s skill document as external weights, applying a rollout‑reflect‑edit‑gate training loop with textual learning rates and rejected‑edit buffers, enabling self‑evolving skills that achieve optimal or tied‑optimal results across 52 model‑benchmark‑environment combinations.
AI agentsMicrosoftSelf-Evolving Skills
0 likes · 12 min read


SuanNiMay 27, 2026 · Artificial Intelligence
SkillOpt treats an agent’s Skill document as a trainable external state, applying classic deep‑learning tools such as epochs, batch size, learning rate and validation gating, and in experiments across 52 benchmark units it lifts GPT‑5.5 performance by an average of 23.5 points while enabling cross‑model and cross‑environment transfer with no additional inference cost.
Agent SkillCross-Model TransferDeep Learning Optimization
0 likes · 11 min read

Microsoft Research introduces SkillOpt, a method that freezes large‑model weights and instead trains a natural‑language skill document as the sole learnable parameter, using a rollout‑reflect‑edit‑gate loop, achieving optimal results across 52 benchmark‑model‑environment combinations and demonstrating strong transferability.
LLM agentsParameter-Efficient Fine-TuningSkillOpt
0 likes · 9 min read

The paper introduces Visual Para-Thinker, a parallel‑thinking framework for large‑scale visual‑language models that uses visual‑centered block and scan path partitions, Path‑aware Attention and Learnable Parallel Rotary Position Embedding, and demonstrates consistent gains across counting, visual search, hallucination and grounding benchmarks.
LPRoPEPa-Attentionbenchmark evaluation
0 likes · 11 min read

The paper introduces Native Parallel Reasoner (NPR), a framework that lets language agents generate and maintain multiple reasoning paths using a three‑stage self‑distillation and parallel reinforcement‑learning training paradigm, achieving up to 4.6× speedup and significant accuracy gains across eight reasoning benchmarks.
AI ReasoningLarge Language ModelsNative Parallel Reasoner
0 likes · 18 min read

The paper introduces OFA-MAS, a one‑for‑all multi‑agent system that learns a universal topology designer using task‑aware graph encoding and a Mixture‑of‑Experts generator, achieving superior performance, OOD generalization, robustness, and efficiency across six major benchmarks.
LLMMixture of ExpertsMulti-Agent Systems
0 likes · 14 min read

DECS, a novel training framework introduced by researchers from Fudan, Shanghai Jiao Tong, and the Shanghai AI Lab, theoretically exposes the flaws of length‑penalty rewards and, through token‑level reward decoupling and dynamic batch scheduling, reduces inference token counts by over 50% while improving accuracy across multiple benchmarks.
DECSLarge Language ModelsReward Design
0 likes · 9 min read

Agent-World introduces a universal training arena that automatically mines real‑world data from the internet to build over 1,900 diverse environments and 19,800 tools, then generates long‑horizon tasks through graph‑based and programmatic synthesis, creating a self‑evolving loop where agents are evaluated, diagnosed, and the environment is refined, achieving state‑of‑the‑art results on 23 benchmarks.
AI agentsAgent-Worldbenchmark evaluation
0 likes · 14 min read

The paper introduces HINTS, a two‑stage self‑supervised framework that leverages Friedkin‑Johnsen opinion dynamics to mine latent human‑driven factors from time‑series residuals, integrates them via attention into state‑of‑the‑art predictors, and demonstrates consistent accuracy gains and interpretability across nine benchmark and real‑world datasets.
Attention MechanismFriedkin-Johnsen modelSelf-Supervised Learning
0 likes · 17 min read


SuanNiApr 3, 2026 · Artificial Intelligence
The article presents the GEMS (Agent‑Native Multimodal Generation with Memory and Skills) framework, detailing its multi‑agent loop, hierarchical memory compression, on‑demand skill modules, and extensive benchmark results that show a lightweight 6B model surpassing larger proprietary systems on complex image‑generation tasks.
GEMSImage GenerationMemory Compression
0 likes · 14 min read


SuanNiMar 20, 2026 · Artificial Intelligence
XSKILL introduces a dual‑stream framework that separates task‑level skills stored as Markdown and action‑level experiences stored as JSON, enabling multimodal large language model agents to continuously improve by extracting, summarizing, and reusing knowledge from past trajectories without modifying model parameters, achieving significant gains across visual tool, multimodal search, and integrated benchmarks.
agent frameworkbenchmark evaluationcontinuous learning
0 likes · 12 min read

The paper introduces GLM, a multi‑agent Graph‑CoT framework with an optimized LLM serving architecture that dramatically improves accuracy, reduces token consumption, lowers latency, and increases throughput across diverse domains, as demonstrated by extensive GRBench evaluations.
LLM optimizationbenchmark evaluationgraph reasoning
0 likes · 10 min read


AntTechOct 14, 2025 · Artificial Intelligence
The Ring-1T model, a trillion-parameter AI system released as open source, leverages advanced reinforcement learning techniques, extensive benchmark evaluations, and custom training frameworks to deliver balanced performance across math, code, reasoning, and creative tasks while highlighting current limitations and future development plans.
AI Modelbenchmark evaluationdeep reasoning
0 likes · 8 min read

MiniMax’s newly released open‑source M1 model, built on the Lightning Attention‑enhanced MiniMax‑01 base, delivers up to 1 million token context, achieves near‑state‑of‑the‑art performance on MRCR and other long‑context benchmarks, and showcases impressive multilingual translation, code completion, and creative applications.
Lightning AttentionMiniMaxOpen Source LLM
0 likes · 11 min read

This article explores how a multi‑agent AI system dramatically improves 0‑day vulnerability detection by automating code audit, reducing false positives, and outperforming traditional static analysis tools in large‑scale real‑world benchmarks.
0day vulnerabilityAI agentAutomated Security Testing
0 likes · 9 min read

A new study from the a‑m‑team introduces “Think Twice”, a test‑time multi‑round reasoning technique that, without additional training or model changes, repeatedly prompts large language models to self‑correct, yielding notable accuracy gains across benchmarks such as AIME, MATH‑500, GPQA‑Diamond and LiveCodeBench, while also producing shorter, more confident answers.
LLMMulti-round reasoningTest-time optimization
0 likes · 6 min read

The article introduces the DistilQwen2.5‑R1 series, which leverages a novel knowledge‑distillation pipeline—including CoT data evaluation, improvement, and validation—to transfer deep reasoning abilities from large models like DeepSeek‑R1 to compact models, achieving superior performance across math, code, and scientific benchmarks and providing open‑source checkpoints and deployment guides for practical use.
AI inferenceLarge Language Modelsbenchmark evaluation
0 likes · 17 min read

This report details the design, data synthesis, three‑stage training pipeline, and benchmark evaluation of the open‑source Steiner reasoning model, which aims to emulate OpenAI o1’s inference‑time scaling while highlighting current performance gaps and future research challenges.
Inference ScalingLLMbenchmark evaluation
0 likes · 14 min read

The article provides a detailed technical overview of Gemma 2, covering its decoder‑only transformer design, novel attention mechanisms, logit soft‑capping, RMSNorm, knowledge‑distillation training on trillions of tokens, extensive pre‑training infrastructure, and benchmark evaluations that demonstrate its competitiveness against larger proprietary models.
AIGemma 2benchmark evaluation
0 likes · 14 min read


AntTechApr 17, 2024 · Artificial Intelligence
LLMRG introduces a novel framework that leverages large language models to construct personalized reasoning graphs, integrating chain reasoning, self‑verification, divergent extension, and knowledge‑base self‑improvement, thereby enhancing recommendation accuracy, interpretability, and performance across multiple benchmark datasets without additional user or item information.
AIInterpretabilityLarge Language Models
0 likes · 9 min read
