Articles & Guides

GigaBrain-0.7 proposes a three-system architecture that connects visual-language understanding, world prediction, and robot action. Trained on more than 37,000 hours of heterogeneous embodied data, it targets stronger transfer across tasks and robot embodiments.

AutoSaddler treats agent harness optimization as an offline learning problem. It diagnoses execution traces, generates targeted code-like patches, and validates updates for broader generalization.

CyberFactory is an open-source pipeline for reconstructing public vulnerability artifacts as executable tasks and filtering agent trajectories through programmatic verification. Its resulting model, OpenAegis, reaches 58.1% Pass@1 on CyberGym under the reported evaluation setup.

SecOPD replaces coarse sequence-level safety feedback with token-level signals during defensive fine-tuning. The paper reports a drop in PISmith attack success rate from 94.0% to 9.0% on defended Qwen3.6-27B.

LAION-BVD is an open video resource built from web-discovered URLs, containing about 80 million downloaded videos with a combined duration of 10 million hours. It supports video, audio, and image-text pre-training through scene extraction, synthetic captions, and key-frame mining.

A new survey argues that smart glasses should be understood as first-person intelligence platforms, not merely cameras or displays. Their real test is whether they can sustain a reliable, correctable, and governed loop from perception to action.

A new environment called Station lets agents from different model families choose research directions, run experiments, and build a shared mathematical literature without a scripted pipeline. The paper reports new results on several construction problems and releases dialogues, proofs, and verification code.

Robotics startup Generalist has reportedly reached a $3 billion valuation after raising nearly $200 million in additional capital led by 8VC. The financing extends its $400 million Series B announced in June, bringing the round’s total to about $600 million.

Chris Malone, OpenAI’s former head of data centers, has left the company after an infrastructure reorganization changed his reporting line. The departure adds to a broader wave of senior exits that is raising questions about execution, governance, and OpenAI’s reported path toward an eventual IPO.

EchoWM is an omnimodal world model designed for interactive, enterable media. As users continuously navigate a scene, it jointly generates 720p video, environmental sound, music, and speech while following first-person or third-person camera behavior.

Prime Agent is an open-source harness for evaluating and operating agents on long-horizon tasks. It combines a persistent REPL, cross-trajectory state, recursive subagents, and recovery tools to separate infrastructure failures from model failures.

MobilePA-Bench introduces an interactive benchmark for mobile planning agents, moving beyond screen manipulation and offline API matching. It evaluates tool use, long-horizon planning, memory, skills, and collaboration under realistic runtime constraints.

Thinkingbox evaluates agents by checking whether they produce the correct end state in policy-constrained business workflows, rather than merely generating valid replies or tool calls. Its benchmark exposes a wide gap between occasional success and repeatable execution.

GameXpert-Bench evaluates coding agents across the full game development lifecycle, covering generation, bug repair, and multi-turn optimization. The results show that agents can build playable foundations, but still struggle with proactive debugging, runtime verification, and regression control.

A study of spoken multi-hop question answering finds that richer retrieval structures can preserve higher absolute accuracy while becoming more sensitive to upstream speech-recognition mistakes. Corrupted query entities are the main path through which those errors spread.

TileMix mixes FP16 and INT8 score computation inside a fused dense-attention kernel by routing hardware-aligned tile groups. The approach keeps all legal token interactions while targeting a better long-context prefill trade-off between quality and throughput.

A new EMNLP 2026 Findings paper shows that search-augmented LLM recommenders can be misled by locally altered web content and promote products that do not exist. Reasoning and current defenses do not reliably stop the attack.

A new audit shows that expanding a retrieval index can change answers even when the model, prompt, retrieval policy, and generation controls remain fixed. The proposed Snapshot Compatibility Audit separates update-driven answer churn from ordinary sampling variability.

RIBOSPAN is a 1.61-billion-parameter bidirectional RNA foundation model pretrained with contexts of up to 10,240 nucleotides. Its evaluations suggest that native long-context training is more reliable than simply extending a short-context model at inference time.

WorldToken combines multiview images, proprioception, and task conditioning into one token for each policy timestep, then models the resulting sequence causally. Its experiments point to the importance of target-domain data and long temporal context in robotic imitation learning.

ReWorld separates short-horizon action control from long-term scene memory during training, then combines a bounded KV cache with pose-indexed landmark retrieval at inference. The approach is designed to preserve revisited scenes without allowing context and compute costs to grow without bound.