#Papers & research

Historian Lindquist: How early engineers introduced 'chaos' into AI to simulate human thinking

Historian Benjamin Lindquist pointed out in a new paper that engineers in the 20th century 1950s injected randomness into AI systems to simulate human creative thinking, which explains the origin of AI hallucinations. Early researchers believed that mathematics could remove the order of art and introduced irregularities through neural networks to inspire insights. Today, systems like ChatGPT or Claude suppress this 'chaos' due to real-world consequences, but the core mechanism still exists.

Factor-Aware Predictive Learning is released: a multi-modal large language model representation method for cinematic video generation.

Factor-Aware Predictive Learning (FAPL) enables video diffusion Transformers to utilize multi-modal large language model (MLLM) representations for sparse temporal reasoning and cinematic control, achieving this through staged grounding and optimization. FAPL consistently outperforms baselines on three cinematic video generation tasks (sparse-to-video, factor-aware editing, and multi-shot generation), validating its contributions in MLLM feature utilization, temporal control, and factor coordination.

OpenDS-World: A benchmark for evaluating autonomous deep search agents in real network environments

OpenDS-World introduces a real-time network benchmark to evaluate autonomous multi-step deep search agents, revealing a significant gap between cutting-edge models and human performance. Among 12 cutting-edge models, the best model achieved an accuracy of 39.0%, while 1000 human evaluations reached an accuracy of 21.1%. Analysis shows that research strategies vary among models, and tool coordination and long-term search management remain major challenges.

Deep Fusion Memory: A block-level parameterized memory architecture for large language models

Deep Fusion Memory (DFM) integrates predicted parametric memory into intermediate hidden states through gated cross-attention, reducing retrieval overhead and improving the performance of knowledge-intensive tasks without the need for online retrieval.

Policy-Mask: A Budget-Aware Adaptation Method for Diffusion Features in Discriminative Tasks

Policy-Mask learns a budget-aware diffusion step, network block, and channel selection strategy to extract compact and task-optimal features for discriminative visual tasks. Selecting intermediate features from a frozen diffusion model for discriminative tasks is complex because useful signals are distributed across denoising trajectories, network depth, and channels.

TTR-Compressor reshapes sequence compression technology

EvalResearchBench: Can AI agents autonomously construct feedback for recursive self-improvement?

EvalResearchBench (ERB) is a benchmark for autonomous evaluation research, allowing AI agents to select or synthesize tasks, implement scorers, and revise within a fixed time and API budget, ultimately freezing an executable evaluator. 9 researchers and 13 candidate models were studied, and the evaluators generated scored approximately 75% consistency on 14 target benchmarks, below the 91% upper limit. Publicly designed tasks remain strong baselines, and the evaluator with the lowest execution cost achieved the highest pairwise consistency.

OmniDR-Bench: Video-centric deep report generation benchmark evaluation.

Researchers have introduced the OmniDR-Bench benchmark and evaluation framework for in-depth study of video-centric deep agents, which can synthesize external evidence to generate interactive reports correlated with video moments. This task has direct practical value, helping users locate relevant video segments, verify claims, and reproduce demonstration steps. OmniDR-Bench includes 50 tasks in Chinese and English, covering 19 real-world subcategories, with a total of 6,414 standards, and the best system score is 63.3.

Cola2: Adaptive Hierarchical Language Model improves hierarchical language generation through continuous semantic compression and balanced discrete generation.

Cola2, through dynamic span splitting, masked diffusion decoding, and stream matching priors jointly optimizes continuous semantic compression, reconstruction-generation balance, and encoder-decoder adaptation, verifying its core design on five downstream tasks. Using 2100 billion training tokens, it demonstrates competitive non-saturated scalability, with a 3460 billion source token checkpoint showing an average of 60.10 points on eight tasks and surpassing Llama-3-8B on HellaSwag and StoryCloze.

RouteWeaver: Unifying LLM routing model selection and execution

RouteWeaver unifies heterogeneous LLM routing models through collaborative training and progressive exploration, improving accuracy in various reasoning tasks; its average accuracy reaches 61.4%, a relative improvement of 11.03% compared to the strongest baseline, and surpasses all comparison baselines in five domains.

Different tasks, different time steps: temporal task-aware quantization for diffusion language models T²Quant

T²Quant framework predicts accuracy sensitivity through minimal unlabeled calibration prompts, allocating activation precision within a fixed bit-width budget, improving existing quantization methods by 6.7 percentage points across 3 diffusion language models and 6 benchmark tests without increasing average accuracy budget.

RLSplat: A feedforward 3D Gaussian splatting method with linear resource budget control.

RLSplat achieves linear reduction in latency, memory, and model size for feedforward 3D Gaussian splatting by assigning Gaussians through regional allocation and using local attention decoding. In zero-shot evaluation, RLSplat outperforms previous budget-controlled methods under all budgets, surpassing the strongest baseline even with a 4x reduction in Gaussian count, and achieves a 4.1x speedup and a 9.6x reduction in peak memory at low budgets.

Agent Mosaic: Identifying confidential leakage risks in agent-based LLM systems.

Agent Mosaic quantified 11,240 responses from 3 LLM agents in simulated and open-source environments, finding that GPT-5.6-Sol had 7.08% of responses (95% CI) in simulations.

Heartbeat thought: Speculative reasoning and tool execution in real-time voice agents.

Heartbeat Thinking enhances the coverage and response speed of voice agents by continuing to prepare tool calls during voice updates. This technology allows voice agents to start preparing tool calls while the user is speaking, addressing the limitation of traditional methods where calls are interrupted by voice updates. In the FDBench v3 benchmark test, Heartbeat Thinking increased tool coverage by 16 percentage points and reduced average final response latency by 36.4%.

Microsoft and Hugging Face use ThinkingBox database validation for RL reward training agents

Microsoft and Hugging Face release ThinkingBox-Training, directly using ThinkingBox's deterministic verification of database final states as reinforcement learning rewards, training Qwen3.8-27B for 50 steps with GRPO

Microsoft releases medical diagnostic system MAI-DxO and step-by-step diagnostic benchmark SDBench

Microsoft's Health MAI team released the AI diagnostic system MAI-DxO and its accompanying benchmark SDBench, with the results published in Science. On 304 NEJM complex cases, MAI-DxO achieved an 83% diagnostic accuracy rate, with doctors achieving 36% with external information and 22% without; on MIMIC-IV-ED's emergency common cases, it was 80%, with doctors at 47%.

AgenticRoboRL framework enhances robot planner performance through post-training

AgenticRoboRL framework improves robot planner performance via supervised fine-tuning and multi-round reinforcement learning, boosting the success rate of a 9B parameter planner from 18.1% to 71.0% on LIBERO-PRO without external memory, approaching Codex (GPT-5.5)'s 72.1% with external memory. The framework adapts to different execution mechanisms, linking recorded planner interactions with task outcomes to achieve a universal reinforcement learning training process through adapters.

Quant-Arena: A large language model agent for benchmark quantitative trading research

Quant-Arena evaluates whether language models can construct profitable trading strategies based on historical A-share data and verifies their performance on closed future year data, as a benchmark quantitative research. The platform allows models to build feature and stock rating code, based solely on feedback from 2022-2023, freeze strategies, and evaluate in 2024. GPT-6 Astra scored the highest among nine cutting-edge models, but no model surpassed the 2024 equal-weight market performance. Quant-Arena is open to all models or agents.

Developer Guide: OmniParseBench's leaderboard benchmark and OCR tests

OmniParseBench Rates 27 Parsers, Using 16,288 Small Tests to Evaluate Their Output, Covering Scanned Forms, Handwritten, arXiv Mathematical Formulas, and Dense Tables. Authors Discovered Through Manual Page Testing That the Top Two Parsers Showed No Statistical Difference, the Rankings Changed Frequently Across Different Document Types, and One Published Data Did Not Match the Actual Files.

How Rescale uses Strands Agents and AI-DLC to build intelligent AI routing agents

Rescale utilizes Strands Agents and the AI-DLC framework to set up intelligent routing agents on AWS in five days and deploy them in production within weeks. The agent classifies end-to-end test failures at night by calling Claude Opus 4.8, retrieves historical failure patterns from the knowledge base, queries MCP logs for runtime evidence, and generates remediation reports containing file references, function names, and recommended fixes for engineers to review.

Redwood Research conducted red team testing on OpenAI's internal agent monitoring system

Redwood Research and METR employees conducted in-depth testing on OpenAI's internal agent monitoring system from July 21 to August 11, finding that approximately 75-94% of GPT-5.5 and above model Agentic Responses API traffic can be monitored.

LASAGNA evaluated the composition of retrieval augmented generation (RAG) pipelines

LASAGNA evaluated the performance of 450 RAG pipeline configurations on 8 benchmarks and 5K-200K document corpora, finding that individual technology advantages at different stages do not necessarily combine effectively, retrieval quality and answer faithfulness optimization directions are different, the best combination varies with corpus size, indicating that isolated test results are unreliable, new RAG technology needs to be evaluated in a complete environment.

𝜏-Entity: How does the speech intelligence agent accurately capture entity information

The 𝜏-Entity benchmark speech intelligence agent captures the accuracy of entity information such as names and phone numbers in 200 tasks. GPT Live 1 has a correct rate of 60.3% under normal conditions, but only 34.0% under noise and voice interference; recommended verification strategies can increase success rates by 14-31 percentage points, but call duration increases by 21-28 seconds.

AKV introduces caching project management for LLM intelligent entity key-value caching

AKV proposes a caching engineering solution, managing LLM agent key-value caches through drop and repos operations, enhancing efficiency and accuracy. Experiments show that tool response rolling drop caching strategy improves the average benchmark score by 1.7 points, reduces costs by 27.7%, maximizes output throughput by 1.2 times, and accelerates P95 request latency by 1.5 times.

Future Chunk Prediction (FCP) enhances the physical plausibility of autoregressive video generation.

Future Chunk Prediction (FCP) improves autoregressive video diffusion models by introducing future chunk supervision, generating videos that better adhere to physical laws and reducing training costs. This method, under the teacher forcing framework, predicts temporally separated video chunks without relying on intermediate visual context, directly utilizing training videos for supervision, requiring no additional pre-trained visual encoder, and not altering autoregressive inference.

Socratic Skill Learning: Enhancing transferable intelligent agent skills through alternating teaching and practice

Socratic Skill Learning improves transferable agent skills through an alternating teaching and practice approach, achieving an effectiveness of 0.801 on the main test set, surpassing trajectory distillation and contrastive distillation methods.

StreamEval: Benchmark for evaluating interactive streaming video generation models

StreamEval benchmarked the performance of interactive streaming video generation models on five tasks, finding reliable request execution and sustained state consistency to be major challenges. The benchmark includes 300 scripted test cases, covering character dialogue, character performance, world generation, instruction editing, and reference editing tasks, scored across four evaluation dimensions (task execution, state consistency, visual fidelity, and audiovisual coordination).

LGWM: A decoder-free, action-conditioned world model for verifying GUI conversion.

LGWM predicts the representation of the next screen to verify GUI conversion, trained on RSWT-BENCH without semantic annotation, achieving 0.987 AUC, decision time of 17 ms, superior to generative GUI world models by three orders of magnitude.

Adobe Research and UCLA jointly propose the Tev model, an open-source 27B model that matches the performance of GPT-6 Astra and Claude Opus 5 on long-range reasoning tasks.

The Tev model proposed by Adobe Research and UCLA matches the performance of GPT-6 Astra and Claude Opus 5 on long-range reasoning tasks. This model is an open-source 27B parameter model that enhances agent capabilities by constructing a 'world of answers' and designing long-range questions.