Top items
- MiniMax Sparse Attention (MSA) cuts per-token attention compute by 28.4× at 1M context, powering the newly released MiniMax-M3 multimodal model.
- EvoArena/EvoMem: a new benchmark and patch-based memory paradigm showing LLM agents struggle (39.6% accuracy) in dynamic, evolving environments.
- WeaveBench: a long-horizon hybrid-interface benchmark for computer-use agents where the best system reaches only 41.2% pass rate.
- SpatialClaw: a training-free, code-as-action framework lifting VLM spatial-reasoning accuracy to 59.9% across 20 benchmarks.
Research papers
MiniMax Sparse Attention (MSA): efficient ultra-long-context attention
Ultra-long-context capability has become indispensable for frontier LLMs — agentic workflows, repository-scale code reasoning, and persistent memory all require jointly attending over hundreds of thousands to millions of tokens — but the quadratic cost of softmax attention makes this untenable at deployment scale. MiniMax Sparse Attention (MSA) is a blockwise sparse attention built on top of Grouped Query Attention (GQA). It works in two stages: a lightweight Index Branch scores key-value blocks and independently selects a Top-k subset for each GQA group (enabling group-specific sparse retrieval while keeping efficient block-level execution), then a Main Branch performs exact block-sparse attention over only the selected blocks. The design emphasizes simplicity and scalability, deliberately streamlined to deploy efficiently across a broad range of GPUs. To turn sparsity into real speedups, MSA is co-designed with a GPU execution path that uses exp-free Top-k selection and KV-outer sparse attention to improve tensor-core utilization under block-granular access. On a 109B-parameter model with native multimodal training, MSA matches GQA in quality while reducing per-token attention compute by 28.4× at 1M context; paired with the co-designed kernel it achieves 14.2× prefill and 7.6× decoding wall-clock speedups on H800. The inference kernel is open-sourced at github.com/MiniMax-AI/MSA, and a production-grade natively multimodal model powered by MSA, MiniMax-M3, has been publicly released at huggingface.co/MiniMaxAI/MiniMax-M3. (108 upvotes; paper)
EvoArena & EvoMem: tracking memory evolution for robust agents in dynamic environments
Most LLM-agent evaluations assume static environments, but real-world deployment is inherently dynamic — agents must continually align their knowledge, skills, and behavior with changing conditions. To close this gap, EvoArena is a benchmark suite that models environment changes as sequences of progressive updates across three domains: terminal, software, and social. Alongside it, EvoMem is a patch-based memory paradigm that records memory evolution as structured update histories, letting agents reason about how the environment has changed through changes in their own memory. Experiments show current agents struggle badly on EvoArena, achieving only 39.6% average accuracy across the evolving terminal, software, and social-preference domains. EvoMem consistently helps: an average gain of 1.5% on EvoArena, plus improvements on standard benchmarks GAIA (+6.1%) and LoCoMo (+4.8%). It also raises chain-level accuracy by 3.7% on EvoArena, where success requires completing a consecutive sequence of related evolutionary subtasks. Mechanistic analysis indicates EvoMem improves evidence capture in memory, better preserving complete evolving environment states. The authors argue that modeling evolution — in both evaluation and memory — is essential for reliable agent deployment. (121 upvotes; paper)
WeaveBench: long-horizon, real-world benchmark for computer-use agents with hybrid interfaces
Computer-use agents (CUAs) increasingly operate in runtimes combining visual desktop control, command-line execution, code editing, browsers, and external tools, yet existing benchmarks tend to test these interfaces as separable capabilities, leaving long-horizon cross-interface orchestration under-tested. WeaveBench is a long-horizon hybrid-interface benchmark with 114 tasks across 8 real-world work domains, grounded in real user requests and publicly verifiable artifacts. Each task requires an agent to combine GUI observations/actions with CLI/code operations within a single trajectory, evaluated on a real Ubuntu desktop inside deployed CLI-agent runtimes augmented with a minimal desktop-control plugin. The authors also introduce a trajectory-aware judge that inspects deliverables, files, screenshots, logs, and action traces — and detects shortcut behaviors such as fabricated visual evidence or hard-coded metrics. Across frontier model-runtime pairings, the best PassRate reaches only 41.2%, showing the benchmark is far from saturated. Critically, the trajectory-aware judge reveals that outcome-only grading substantially overestimates agent performance. WeaveBench exposes a key gap in CUA evaluation and provides a testbed for whether agents can orchestrate GUI, CLI, and code operations across long-horizon real-world tasks. (95 upvotes; paper)
SpatialClaw: code as the action interface for agentic spatial reasoning
Spatial reasoning — determining where objects are, how they relate, and how they move in 3D — remains a fundamental challenge for vision-language models (VLMs). Tool-augmented agents try to address this by adding specialist perception modules, but their effectiveness is bounded by the action interface through which tools are invoked. Existing spatial agents either use single-pass code execution (committing to a full strategy before observing any intermediate result) or a structured tool-call interface (which limits flexible composition and per-task tailoring) — both constrain open-ended, complex 3D/4D reasoning. SpatialClaw is a training-free framework that adopts code as the action interface. It maintains a stateful Python kernel pre-loaded with input frames and a suite of perception and geometry primitives, letting a VLM-backed agent write one executable cell per step conditioned on all prior outputs. This lets the agent flexibly compose and manipulate perception results and adapt its analysis to intermediate text and visual observations and to each problem’s demands. Evaluated across 20 spatial-reasoning benchmarks spanning static and dynamic 3D/4D tasks, SpatialClaw achieves 59.9% average accuracy, outperforming the recent state-of-the-art spatial agent by +11.2 points, with consistent gains across six VLM backbones from two model families — and no benchmark- or model-specific adaptation. (83 upvotes; paper)