Thirteen Skills Push Hallucination Detection to 92%

Featured

01 Agent | Thirteen Skills Make Research Auditable

Generating paper-shaped text is easy. Making every claim withstand scrutiny is much harder. Spark-to-Paper breaks research into 13 composable skills that run inside existing coding assistants.

The skills cover literature retrieval, experiment execution, evidence-based claim revision, and editable figures. The system defines required evidence before running experiments. It then observes results and revises or abandons claims instead of selecting favorable data afterward.

Deterministic completeness checks work alongside self-review. Limits prevent repeated experiments from creating endless self-refutation loops. Across eight controlled topics, the full review stack raised hallucination detection from 14% to 92%. It also achieved 99.5% citation validity and 96.4% figure editability.

Each paper used 11.9 million tokens, cost $8.10, and took 3.2 hours on average. That looks more like an operational research pipeline than an expensive demo. Still, eight topics leave open questions about metric definitions, task difficulty, and failure cases.

Key takeaways:

• Evaluate research agents through their evidence trails and ability to catch local errors.
• Define the evidence an experiment must produce before seeing its results.
• The 13 skills offer a lightweight architecture for research workflows inside existing coding assistants.

Source: Spark-to-Paper: End-to-End Research Paper Generation as a Composable Skill

02 Reasoning | Strong Harnesses Lift Weak Models

AI4AI raised weak-model performance from 0.49 to 0.91 without retraining. A strong model instead builds an execution harness before testing begins.

Using 5% of the data, the strong model repeatedly debugs the workflow. It moves unstable reasoning into deterministic code, task routing, and strict answer formats. The weak model then executes within that structure.

Most gains came from redesigning the reasoning environment, not giving the weak model more time to think. One question remains: does the harness capture reusable task structure, or solve most reasoning upfront? Results currently cover only four theory-of-mind benchmarks.

Cost comparisons must include harness construction, validation, and maintenance. Comparing only the weak model's per-call price hides much of the system cost.

Key takeaways:

• Model improvements do not always require parameter updates. Start by improving the reasoning environment.
• Separate reusable task structure from benchmark-specific reasoning performed upfront.
• Include strong-model development and maintenance when calculating total cost.

Source: AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses

03 Safety | Adversarial Tool Environments at Scale

Agent safety testing usually relies on handcrafted environments with a few planted prompt injections. ToolHazard instead generates stateful, executable adversarial tool environments in batches.

An environment simulator, attacker agent, and user simulator find viable injection points. They create environment-specific payloads and build long-running tasks tied to internal state. Teams could add this process to continuous testing across tools, permissions, and injection locations.

The reported experiments show that injection timing and location affect attack success. Generated alignment data improved safety on two benchmarks without reducing normal task performance. Yet synthetic volume says little about whether scenarios match real business workflows.

Key takeaways:

• Add adversarial environment generation to continuous testing instead of maintaining only fixed attack samples.
• Cover tools, permissions, injection locations, and timing in the test matrix.
• Evaluate scenario volume separately from real-world business relevance.

Source: ToolHazard: Scaling Adversarial Environments for Security Evaluation and Alignment of LLM-based Agents

04 Multimodal | StateFlow Grounds Cameras in 3D

StateFlow stores scene elements and camera configurations in a persistent, structured 3D state. Its previsualization workflow has three stages: building, updating, and accessing the world.

The build stage lifts 2D content into a coherent 3D world. Updates preserve world memory while converting user intent into structured state changes. During access, render feedback helps correct camera plans into visually feasible trajectories instead of relying only on VLM semantics.

Final visual quality is a separate concern. StateFlow still uses existing video models to improve its output, so polished results do not directly measure the state layer. Teams should test world consistency across repeated changes and whether render feedback reliably improves camera motion.

Key takeaways:

• Evaluate world-state updates, camera access, and final visual quality separately.
• World memory determines whether scenes stay consistent across multiple revisions.
• Render feedback can correct camera plans into visually feasible trajectories.

Source: StateFlow: Building, Evolving, and Accessing 3D World States for Previsualization