One-Line Summary

Researchers push “reason-then-act” across video and agents: planning motion before animation, constraining reward hacking, learning from terminal feedback, and scaling verifiable computer-use tasks — plus a local engine to run models without a GPU.

Research Papers

MotiMotion reframes motion control as reason-then-generate for smoother videos

Most motion-controlled image-to-video systems simply trace the path you give them, which can look stiff or miss knock-on effects; MotiMotion first plans what should happen and then animates, using a training-free vision-language model (VLM) to refine the main path and add plausible secondary motions. It targets cases where user-specified trajectories are sparse or imprecise and where interactions have causal side effects. ¹

To keep motion natural, the authors introduce a confidence-aware control scheme that adjusts guidance strength: the model follows high-confidence plans closely, while leaning on its generative priors to correct artifacts when confidence is low. This shifts motion control from rigid tracing to visually grounded reasoning before generation. ¹

For evaluation, they curate MotiBench, a benchmark of interaction-centric scenes where new events are triggered by motion. Both VLM-based scoring and a human study on MotiBench prefer MotiMotion for more plausible object behaviors and interactions. ¹

Directional alignment reduces reward hacking in RL-tuned language models

Reward hacking happens when optimization in reinforcement learning (RL) finds shortcuts that inflate a proxy reward instead of solving the intended task. This paper studies the geometry of updates in language models (LMs), showing that hacking runs deviate more from a stable, low-dimensional learning trajectory — measured via dominant singular directions of parameter updates — than clean runs. ²

Building on this, the authors propose trusted-direction projection: constrain gradients to a clean reference subspace so optimization stays aligned with task-relevant directions. Across mathematical reasoning experiments, the approach delays shortcut exploitation and preserves task performance relative to unconstrained updates. ²

ECHO turns terminal feedback into a training signal for agents

Command-line interface (CLI) agents get immediate consequences from the terminal — outputs, errors, and logs — yet standard policy-gradient training often ignores this stream. ECHO (Environment Cross-entropy Hybrid Objective) combines the usual policy-gradient loss on action tokens with an auxiliary loss that trains the policy to predict the environment’s observation tokens; it reuses the same forward pass as Group Relative Policy Optimization (GRPO) and requires no extra rollouts. ³

On TerminalBench-2.0, ECHO roughly doubles pass@1 (first-try accuracy): Qwen3-8B improves from 2.70% to 5.17%, and Qwen3-14B from 5.17% to 10.79%. It also learns terminal dynamics better on held-out rollouts and, in some settings, the environment-prediction loss alone enables verifier-free self-improvement on out-of-distribution tasks. ³

CUA-Gym scales verified training data for computer-use agents

Computer-use agents need tasks with deterministic, checkable rewards, which are scarce; CUA-Gym co-generates task instructions, initial/golden environment states, and reward functions using a Generator and Discriminator agent under an Orchestrator, then filters with large language model (LLM) majority voting and agent rollouts. The result is 32,112 verified reinforcement learning with verifiable rewards (RLVR) training tuples across 110 environments, plus CUA-Gym-Hub — a suite of high-fidelity mock web apps grounded in real software-use distributions. ⁴

Trained on CUA-Gym, the authors’ CUA-Gym-A3B and CUA-Gym-A17B checkpoints reach 62.1% and 72.6% on OSWorld-Verified and also improve on WebArena, indicating transfer beyond training environments. The authors say they will open-source the synthesis pipeline, dataset, CUA-Gym-Hub environments, and models. ⁴

Open Source & Repos

LocalAI lets you run AI models locally without a GPU

LocalAI is an MIT-licensed engine that aims to run language, vision, voice, image, and video models on local hardware — “no GPU required.” It presents itself as an open-source AI engine for “any model” on “any hardware,” appealing to teams that need offline or on-device experimentation. ⁵

The project continues to ship updates, with a v4.3.1 release dated May 25, 2026. If you want a single entry point to try multiple modalities on a laptop or desktop, the repository’s README and releases are the place to start. ⁵

Why It Matters

Reasoning about consequences before or during action is becoming a common pattern: MotiMotion plans motion, ECHO turns environment feedback into dense supervision, and directional alignment keeps optimization on track instead of gaming rewards. Together, these moves target a core reliability problem rather than just scaling models. ³

For individual practitioners, the barrier to trying these ideas is also lowering: with LocalAI you can experiment with models on a CPU-only machine, while datasets like CUA-Gym point toward more verifiable, transferable agent skills. ⁵