▲ 325 · 16h agoBuilding better training examples for AI assistants that use tools
When an AI assistant completes a task successfully, its step-by-step record isn't always good teaching material. This work deliberately designs new practice scenarios that fill in missing skills, then trains assistants only on the steps that were justified by what they could actually see at the time. It matters because better-designed practice, not just more practice, could make AI helpers noticeably more reliable.
Technical summary+
Trajectory2Gym: Decision-Grounded Compositional Interaction Synthesis for Tool-Use Agents
Trajectory2Gym synthesizes executable interactions targeting under-supported compositional structures, supervising only decision-grounded actions with parent-normalized SFT weighting. Across five Qwen3 scales it improves Tau1, Tau2, and BFCL at every scale, raising the equal-weight three-benchmark mean by 16.9–29.8 points, with the largest gains on stateful multi-turn tasks.
A video AI that notices what you want before you ask
Most video AI helpers only respond when you explicitly ask a question, missing hints about what you actually want and failing to keep up as your goals change. This team built a large collection of test examples covering spoken, unspoken, and shifting intents, plus a video model that watches continuously and chimes in at the right moment. This could make video assistants feel much more like an attentive companion than a question-answering machine.
Technical summary+
StreamVibe: Proactive Video Agents for Evolving User Intent
StreamVibe-Bench provides 4.1K eval and 85K training samples spanning explicit, implicit, and dynamic intents; StreamVibe-VL infers evolving intents from streaming observations, with StreamVibe-Harness tracking intent as persistent state via hierarchical memory. The model retains general video understanding while improving response timing and answer quality by 10.7% over baselines.
An AI team leader that invents its own way to delegate work
AI systems made of multiple helper programs usually follow a fixed plan for who talks to whom. This work lets the main AI decide when it needs help and write its own custom coordination recipe on the fly, calling helpers in sequence or in parallel as needed. It matters because flexible teamwork, rather than rigid teamwork, could let AI handle a much wider range of hard problems.
Technical summary+
DynaFlow: Learning Free-Form Multi-Agent Orchestration
DynaFlow injects free-form executable workflow synthesis into the main agent's action space, training only the main agent via SFT on distilled trajectories plus filtering-based refinement, with sub-agents fixed. It reaches 88.08% overall accuracy in Forced mode across six benchmarks, beating the strongest baseline by 10.90 points with the same Qwen3.5-27B backbone.
Stress-testing the scaffolding around AI agents, not just the AI
An AI assistant's success depends not only on the model itself but also on the supporting framework—the code that handles its memory, tools, and checklists—yet most tests ignore that framework. This work created 95 expert-level tasks and carefully recorded every step of thousands of agent runs to see where the scaffolding, not the model, fails. This could help developers build more dependable AI systems by pinpointing exactly which part of the setup broke.
Technical summary+
Trial by Trace: Benchmarking AI Agent Harnesses for Reliability and Failure Attribution
TraceTrial-Bench evaluates harness reliability across seven components using 95 mechanism-oriented tasks, with Trajectory Tracing, Multi-source Evidence Acquisition, and Failure Attribution modules. Across 4 harnesses, 9 models, and 3,420 runs, harness scores ranged 59.8–72.0% while only 31.1% of runs succeeded.
A cheaper way for AI to teach itself while it learns
AI models improve by practicing and checking their own answers, but the usual technique of comparing against a separate 'teacher' version of itself is unstable and memory-hungry. This method instead keeps a small set of snapshots of the model's earlier decisions as fixed reference points, making self-teaching steadier and cheaper. It matters because it could make advanced reasoning training faster and more affordable.
Technical summary+
Function-Space Temporal Self-Distillation for Policy Optimization
FSPO is a single-policy method storing sparse historical next-token distributions as fixed temporal targets, adding privileged disagreement correction, reliability-adaptive grounding, and quality/coverage-aware temporal memory. Under a 5-hour budget it improves average accuracy by ~4 points over the strongest baseline on reasoning and tool-use benchmarks, cutting peak memory from ~70 GB to 56 GB on Olmo3-7B-Instruct.
A practice world where AI assistants coach simulated people
Researchers built a virtual world where a smart assistant guides a simulated human who looks around, listens, and acts on their own, just like a real person. This makes it possible to test whether AI helpers can truly assist people with everyday physical tasks, instead of only robots.
Technical summary+
InteractiveWorld: A Verifiable Human-Centric Environment for Building Better Interactive MLLM Assistants
InteractiveWorld is a closed-loop, human-agent simulator where an MLLM assistant instructs a human that independently observes, understands, and acts, with state-based verifiable rewards; InteractiveBench provides 120 multi-turn tasks across six categories for evaluating egocentric assistants.
A blueprint for AI systems that do real work
This paper offers a big-picture framework for how AI agents that perform real jobs should be built and judged, treating them as systems made of a powerful model plus supporting software. It aims to guide researchers toward agents that finish delegated tasks reliably, with results you can verify.
Technical summary+
Work Agents: Systems, Capabilities, and Optimization
The paper frames work agents as foundation models plus runtime harnesses operating in stateful workspaces, introducing work persistence/openness dimensions, five work forms, five core capability families, and optimization strategies spanning model, harness, and joint tuning, plus outcome-focused evaluation methods.
Videos of characters that stay consistent from any camera angle
When a camera swings far around a character in a generated video, the character's look often gets distorted or changes identity. The researchers created a large dataset plus a method that uses 3D geometry as a guide, keeping characters consistent even during big camera moves.
Technical summary+
WorldCast: World-Consistent Character Video Generation with Large-Angle Camera Control
WorldCast back-projects the initial frame into a depth-based 3D point cloud rendered along the target camera trajectory for geometric guidance and coverage masks, and uses Geometry-Aligned Memory Attention with multi-view anchors to recover unseen regions; they also release CLACT, a large-scale benchmark for large-angle camera trajectories.
Generating and editing 3D sound by following a plan
This system can create and edit immersive, three-dimensional audio, like placing voices and music at specific spots around a listener. It works by first writing a 'plan' of what sounds should do, then rendering that plan into realistic spatial sound, and can also edit existing audio while preserving everything you didn't ask to change.
Technical summary+
Spatial Audio Generation and Editing with Executable Plans and Self-Distilling Transfusion
AMBIT uses a shared Transfusion backbone coupling autoregressive planning with rectified-flow rendering over executable FOA ScenePlans, an FOA-aware continuous VAE with spatial covariance matching, and on-policy self-distillation where target-conditioned teachers supervise student-generated planning and rendering; it outperforms baselines on generation quality, edit fidelity, and unedited-source preservation.
A test of whether AI agents learn from experience over time
Instead of testing AI on one-off puzzles, this benchmark checks whether agents genuinely improve as they accumulate experience across long, realistic work assignments. Even the best AI systems passed less than half of the challenge levels, showing that lasting self-improvement remains hard for today's models.
Technical summary+
Continual Learning Exam: Benchmarking Long-Horizon RSI and Experiential Learning in LLM Agents
CLE is a 50+ expert-built benchmark with 120 human-authored tasks organized as five-level ladders (600 levels, 89 roles, 12 domains) graded by rubric-based evaluation agents; across 17 harness-model configurations, feedback plus one revision yields a median +5.3pp completion gain, but the best system clears only 47.7% of levels and all five levels in 23.3% of tasks.
The most-upvoted papers on Hugging Face Daily Papers from the last few days. Explanations are written by The Wire from each paper's abstract.