Learning Workflows Lift Design Agents Across Seven Setups

Featured

01 Recursive Feedback Lifts Every Design Setup

AutoDesign turns the static DesignHarness into a workflow that improves through output feedback. A meta-optimizer reviews each rollout and guides code agents as they recursively revise the harness.

With the learned DesignHarness, all seven controlled code-agent and model setups improved. Their average score rose from 54.99 to 67.39. AutoDesign scored 78.32 on PosterBench, covering 100 papers across five disciplines. That beat the closed-source Claude Design system by 7.45 points.

These results do not prove that workflows generally matter more than models. They do show consistent gains across configurations within paper-to-poster generation. One autonomous run completed 253 tool calls and 11 editing rounds in 40 minutes for under $3. Human reviewers rated the output as average conference-poster quality, and it ranked first in blind preference tests.

Key takeaways:

• Treating the workflow as a learnable object produced consistent gains across controlled configurations.
• Evaluate design agents by stable improvements across model setups, not one exceptional example.
• Test recursive feedback first on long workflows with clear goals and stable evaluation criteria.

Source: AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design

02 Persistent Worlds Need External Memory

Long-running world models usually trade memory capacity for response speed. Evoke reframes that constraint as a systems architecture problem.

It stores scene geometry in an external state database indexed by camera position. Each generation retrieves only the information needed for the current view. The generator's context therefore stays bounded as interaction time grows.

The teacher receives long-horizon supervision through chunking, distant-frame retrieval, and a global linear-attention state. A 30-second distribution-matching objective then trains a three-step student on self-generated rollouts. This transfers resistance to long-term drift while preserving responsive conditional control.

Persistent memory, long-horizon supervision, and low-latency generation can now scale as separate modules. A 1.5-second clip still takes 2.11 seconds on one H200, so strict real-time generation is still out of reach. Teams building simulations, games, or embodied agents should focus on state-generation separation, not just maximum video length.

Key takeaways:

• Externalize persistent state so longer sessions do not directly increase generation cost.
• Few-step students depend on teachers with genuine long-horizon supervision, not just strong short-clip generation.
• Evaluate state retrieval, long-term drift, and end-to-end latency alongside video duration.

Source: Alaya-EVOKE: From Linear-Scaling Supervision to Endless World

03 Correct Facts Can Still Produce Wrong Answers

A model may judge every atomic claim correctly and still fail when options combine them with AND, OR, or NEITHER/NOR. The failure may lie in answer composition rather than knowledge.

This method removes compound options from the model's job. It splits each option into atomic judgments and scores opposing hypotheses. An integer-programming solver then assembles the final answer under explicit logical constraints.

That division is reusable: models handle uncertain semantic judgments, while solvers enforce deterministic consistency. Macro-F1 rose from 48.3 to 77.0 on one dataset and from 47.0 to 75.6 on another. NEITHER/NOR cases gained the most. Performance still depends on accurate decomposition and well-calibrated scores, which need broader testing.

Key takeaways:

• Check answer representation and composition before blaming failures on limited model capability.
• Give uncertain semantic judgments to the model and hard logical constraints to a conventional solver.
• Test whether decomposition errors or poor calibration become amplified during composition.

Source: From Atomic Evidence to Logical Composition: Structured Compositional Reasoning over Compound Answer Options

04 Does One Pipeline Make a Scientific Foundation Model?

Scientific agents are moving beyond papers and tool calls. They now process documents, images, and time series while advancing through long task chains.

Intern-S2-Preview's main contribution is not simply its 397B parameters. It combines scientific multimodal pretraining, supervised fine-tuning, multitask reinforcement learning, and online distillation in one pipeline. Partial rollouts with off-policy correction and length adjustment help stabilize training.

The architecture adds numerical forecasting for time series and a separate 4B memory decoder. This decoder adapts the frozen 397B backbone to specific fields. On biology instruction tasks, memory expansion raised the average score from 56.92 to 60.32.

That supports low-cost specialization without modifying the backbone. A transferable scientific foundation still needs inexpensive adaptation to disciplines excluded from training. Otherwise, this is closer to several scientific workflows consolidated into one large engineering system. Cross-domain transfer experiments are needed to settle the claim.

Key takeaways:

• Judge scientific foundation models by adaptation cost in unfamiliar disciplines, not only aggregate benchmarks.
• A frozen backbone with a small external memory module may offer a practical route to cheaper domain adaptation.
• A unified training pipeline demonstrates engineering integration, not general scientific reasoning by itself.

Source: Intern-S2-Preview: Scientific Agentic Foundation Model