Benchmark

Meta's Muse Spark Open is an 816B-parameter open-weight multimodal model: Meta’s Muse Spark Open is a new open-weight model with 816 billion parameters that processes text, images, video, and audio with up to one million tokens of context. It performs roughly on par with the most capable open-weight Chinese AI models on benchmarks and is optimized to run locally on Apple's M5 Ultra.

Meta • The Decoder

ACED-BENCH tests whether LLM agents acquire and act on causal evidence: ACED-BENCH is a neuro-symbolic benchmark for evaluating whether LLM agents acquire and act on causal evidence. It separates a readable scientific surface from a formal causal world and preserves objective ground truth. Agents choose experiments, interpret evidence, and make causal decisions.

AI Benchmark • arXiv

LHTB 2.0 expands its benchmark to 77 tasks across 13 domains: LHTB 2.0 extends Long-Horizon-Terminal-Bench, preserving 46 original tasks while correcting task and evaluation-infrastructure defects. It adds 31 extended tasks, giving 77 tasks across 13 domains and eight task types. Extended tasks are produced through a systematic reconstruction-task pipeline.

AI Agent Benchmark • arXiv

Study finds LLM reasoning uses 3.7× more inference energy on average: A study measured inference energy for 86 open-weight LLM checkpoints across seven benchmarks and six hardware platforms. Most configurations used a median 2.6–4.8× more energy than another configuration scoring at least as well, with gaps up to 27×; reasoning raised energy about 3.7× on average.

AI Research • arXiv

Gemini 4 Argon - Out and Tested!!: Sam Witteveen posted a video testing Gemini 4 Argon, comparing it with other models, reviewing its strengths and generated outputs. The description links to a Google blog post introducing Gemini 4 Argon and an Artificial Analysis page but reports no specific benchmark results.

Artificial Intelligence • Sam Witteveen

Tsinghua researchers release SRHarness for formula discovery, MechBench for mechanism recovery: Tsinghua researchers released SRHarness, an open-source runtime for agentic symbolic regression, and MechBench, a benchmark that tests whether AI can recover the mechanisms behind formulas. SRHarness targets formula discovery, while MechBench evaluates mechanism recovery.

AI Research • Pandaily

Google releases Gemini 4 Argon worldwide, saying Argon cut coding tasks from months to weeks: Google released Gemini 4 Argon worldwide via its API, Gemini app, and Google Antigravity. It is aggressively priced but trails Anthropic's Claude Opus 5.5 and Meta's Muse Wonder Full in independent benchmarks. Google says Argon internally cut coding tasks from months to weeks.

AI Models • The Decoder

MagicLab's EchoPolicy-0 tops AI2's MolmoSpaces board with 75.4% manipulation success: MagicLab said its VLM agent framework EchoPolicy-0 averaged 75.4% across four MolmoSpaces tasks, and AI2's published leaderboard data lists it first. Its MagicBrain component reached 94.4% on the pick task.

AI & Robotics • Pandaily