AesCode 8B Gains 31.1 Points in Its Own Benchmark

Featured

01 Code Intelligence: Why Correct Slides Still Look Bad

Code-generated slides and dashboards must satisfy competing goals. Their text and data must be correct, while their layouts and colors must look good. AesCode represents each canvas as a design graph instead of asking the model to copy a reference image directly.

The graph connects elements, attributes, and relationships to both code and rendered output. This makes each score traceable. Deterministic verifiers check properties that can be parsed from code or pages. A vision-language model judges harder visual qualities against sample-specific criteria. Each reward channel is normalized separately before reinforcement learning. Aesthetic scores cannot mask content errors, while strict checks do not reduce design to mechanical templates.

The abstract reports visual-quality gains of 31.1 and 22.4 percentage points for the 8B and 32B versions. Both comparisons use same-size Qwen3-VL baselines with identical reference images. These are internal results, so gains across products, styles, and real editing workflows still need testing.

Key takeaways:

• Separate content correctness from visual quality so one combined score cannot hide failures.
• Use a design graph as a shared interface between code, rendered output, and visual evaluation.
• A 31.1-point gain supports the training approach, but does not establish real-product performance.

Source: AesCode: Aesthetic Code Generation with Decoupled Cross-Modal Rewards

02 Agent: Finishing After Mistakes Is the Scarce Skill

Terminal agents often struggle to recover from command errors or faulty implementations during long-running execution. ReTerm makes those failure states the starting point for training. It begins with supervised fine-tuning on 10,000 recovery demonstrations.

The agent then branches from its own failed states and tries several independent recovery paths. Whether each path eventually completes the task supplies the reinforcement signal. Training Qwen3.5-9B this way raised Pass@1 on Terminal-Bench 2.0 from about 20.2% to 34.8%.

Recovery training clearly raises the capability floor, but nearly two-thirds of tasks still fail. More environments must show whether the skill generalizes. Coding-agent teams should retain failures as training assets containing the error, context, and subsequent correction.

Key takeaways:

• Failed trajectories matter because they can teach agents to finish tasks from erroneous states.
• Branching from failure points turns an agent's own execution history into reinforcement-learning signals.
• A 34.8% Pass@1 is meaningful progress, but cannot support reliability claims.

Source: ReTerm: Recovery Training for Self-Improving Terminal Agents

03 Architecture: Four Constraints for Low-Latency Routing

Jeva replaces text generation with a dedicated scoring interface for structured, low-latency decisions. It separately predicts when no candidate fits, rather than confusing absence with uncertainty among available choices.

Symmetric aggregation keeps decisions unchanged when candidates are reordered. Shared scoring naturally supports candidate sets that grow or shrink. Jeva also stores candidate representations as reusable read-only memory. Each request updates only its own state, while recurring candidate representations and their layer projections can be reused across requests.

The abstract reports no decision flips across the tested candidate permutations. Jeva also scales to thousands of candidates and cuts latency through reuse. Exact throughput, cache costs, and task coverage still require confirmation from the full paper. Low latency may come from stopping text generation altogether, not merely generating faster.

Key takeaways:

• For frequent candidate decisions, evaluate dedicated scoring interfaces instead of only compressing autoregressive decoding.
• Test null answers, candidate reordering, and candidate-set expansion separately before release.
• Systems with frequently reused candidates should measure latency gains from pre-encoding and cross-request caching.

Source: Jeva: System One Models from First Principles