Skip to main content
← AI Knowledge Library
AI Knowledge Library / Research

The daily AI brief.

A considered selection of AI research. What it says, why it matters and where to read the original.

Use with your agent
Add Scaled Memory to your brief.

Copy and paste this into your agent:

Include research from https://www.scaledmemory.com/accelerators/catalogue/ai-knowledge-library in my daily brief. Read the daily brief and source feed linked on that page, using WebMCP if available. Include why the featured stories matter, the original source links and edition date. Flag stale or incomplete content.

Read the research bundle
Subscribe via RSS
Subscribe to Scaled Memory.

The same daily brief and source links, in your RSS reader or agent.

Open RSS feed

The feed mirrors the same brief and 7 am–7 am Melbourne source window.

Page options
Read JSON Agent reading guide

Public, read-only formats for your tools and agents.

12 November 20263 selected storiesCurated by Scaled Memory
Latest available edition: 2026-11-12. Today's update is delayed. Latest edition: 12 November 2026. Displayed edition: 12 November 2026.
  1. TRIAD finds multimodal retrieval struggles with complete, evolving requests

    TRIAD evaluates 17 systems across six multimodal retrieval subsets, each containing 1,000 queries and 1,000 candidates, using multi-positive mAP. Performance generally declines with hierarchical distance and negative constraints; multi-turn retrieval scores lower in 39 of 40 comparisons with paired final intent, though candidate pools highly overlap without being identical. Evaluations are zero-shot and descriptive across public checkpoints; downstream agentic task completion was not tested.

    Why read it

    It separates similarity ranking from satisfying hierarchy, exclusions, preserved conditions, and revised dialogue state.

    Read the original: TRIAD: Multimodal Retrieval across Abstraction, Constraints, and State

    Source publication date unknown.
    Added to the library:

  2. Yuanbao-Reranker reports gains from grounded data and preference distillation

    Yuanbao-Reranker trains 0.6B, 4B, and 8B dense rerankers with document-grounded synthesis, teacher-screened negatives, and bounded preference distillation. On identical Top-100 candidate pools retrieved by Qwen3-Embedding-0.6B, results use nDCG@10 multiplied by 100: five-panel averages were 70.96, 73.63, and 74.21. Panels covered Chinese, multilingual, English, long-document, and code retrieval; results are reported by the authors.

    Why read it

    The evaluation ties reported reranking gains to a fixed first-stage pool, an explicit metric, and diverse language, length, and domain panels.

    Read the original: Yuanbao-Reranker: Document-Grounded Synthesis and Preference Distillation for Text Reranking

    Source publication date unknown.
    Added to the library:

  3. AutoRefine improves code-judge calibration but exposes workflow limits

    AutoRefine uses workflow outcomes and an abstaining teacher to rewrite code-judge prompts. On 85 held-out oracle tasks, including 13 failures, rubric-judge Scott’s pi rose from 0.36 to 0.60; replay results excluded about 5 percent of tasks for timeouts or errors. On 17 independently held-out current diffs, agreement rose from 0.16 to 0.51, while the prior judge remained better on 16 older diffs.

    Why read it

    Results suggest outcome-based calibration can help, but small samples, one internal agent and workflow, distribution shift, and hidden failures constrain generalization.

    Read the original: AutoRefine: Auto-Calibrating Code Judges from Workflow Outcomes

    Source publication date unknown.
    Added to the library:

About this edition

Latest edition: 12 November 2026. Selection: complete. Edition dates describe our selection, not when sources were published.

Selection window: 2026-11-10T20:00:00Z to 2026-11-11T20:00:00Z. Next edition due: 2026-11-12T21:00:00Z.

This editorial selection is not a full-library search or an archive. RSS contains the latest displayed selection; infrequent polling may miss editions.

Checks for updates every five minutes while this page is visible.