November 14, 2026

Radar Weekly, November 7–13, 2026

Tests for action-sensitive world models, evidence-aware retrieval, startup approaches to persistent agents and robot evaluation, and new TMLR records.


This AI-assisted review covers November 7–13, 2026 (UTC), using records visible in Pub-AI’s four Radars. The dates below distinguish when research or a company entered the Radar from when the original source published it.

World Models

The World Models Digest showed 122 papers with Radar dates in the review window. Three preprints stood out for testing whether actions change what the model predicts, using different evaluation metrics.

Preprint

Prediction or Persistence?

Test what the predictor adds beyond reusing the last representation.

The study pairs each video model with a version that reuses its last representation without invoking the predictor. In the author’s household-video retrieval test, the predictor improved top-1 accuracy by 29.1 percentage points for V-JEPA 2.1 and about 3–4 points for V-JEPA 2 on the same trials.

Scope. The task measures retrieval of the true continuation among candidates. The reported gain concerns that evaluation.

Preprint

Does Your World Model Know Actions Have Consequences?

Change the action, then check whether the predicted response changes correctly.

Gaggar and colleagues propose changing a candidate action and comparing the model’s predicted response with the observed response. This gives teams an additional test when evaluating a world model for planning.

Scope. Action-response testing complements reconstruction and task-success metrics; it addresses a different evaluation question.

Preprint

Driving World Models Move the Camera, Not the Scene

Following a camera trajectory can conceal an inaccurate scene response.

Feng and Ao find that a forecast can follow a planned camera trajectory while reproducing only about a third of the corresponding scene response they measured in real video. Their study offers another test to run alongside reconstruction or task success when evaluating a model for planning.

Scope. The comparison concerns the driving-video scene response measured in the authors' study.

Browse the World Models Digest →

AI Techniques

AI Techniques listed 115 discoveries during the window, including older methods. Both examples below were discovered by the Radar on November 10; the records list reproduction as unevaluated.

Research method · Reproduction unevaluated

EviCover

Measure retrieval and answer quality separately.

EviCover selects conversation turns jointly to cover the evidence a question needs. On 108 multi-hop REALTALK questions, the authors report that recall in the final ten-turn context rose from 0.349 to 0.379. This helps locate whether a system lost evidence during retrieval or failed to use it during generation.

Limitation. The paired answer evaluation did not establish a gain in correctness.

Research method · Reproduction unevaluated

AutoEMatic

Adjust quantization as inference runs.

AutoEMatic selects precision for weights, activations and the KV cache as the model runs. The authors report retaining about 90–98% of NVFP4 decode throughput on GB300 GPUs across their evaluated configurations.

Limitation. Transferring the result to another serving stack depends on support for the required formats and kernels.

Browse AI Techniques →

AI Startups

AI Startups listed 29 discoveries, all dated November 10 in the Radar. These two examples approach infrastructure for persistent agent work and robot evaluation.

Company-reported

Superserve

Keep an agent's files and execution state through pauses.

Superserve describes cloud virtual machines that keep an agent’s files and execution state when the machine pauses and resumes. The company entered the Radar on November 10.

Verification. Capabilities are company-reported. The Radar did not confirm a first public launch date or product access.

Company-reported

SE3 Labs

Use failures on real robot tasks to guide data collection.

SE3 Labs describes applying a consistent evaluation protocol to robot policies on real tasks, then using observed failures to decide what data to collect. The company entered the Radar on November 10.

Verification. Capabilities are company-reported. The Radar did not confirm a first public launch date or product access.

Browse AI Startups →

Publication Radar

Publication Radar recorded 74 TMLR publication notices and 13 explicit arXiv revisions first detected during the review window. TMLR lists the 74 records with month-level precision (November 2026). Two examples from the official journal records follow.

Journal record · TMLR

From Text to Voice

Evaluation of tool-calling language-model agents.

The paper appears among the November 2026 TMLR records. The publisher index links code for the paper.

Date precision. The publisher lists a month, not an exact publication day. This report highlights the journal record.

Journal record · TMLR

Naïve Difficulty Conditioning Fails on ShapeNet-55

A matched-control audit, with code linked by the publisher.

The paper appears among the November 2026 TMLR records. Its journal record is one example of the publication notices detected in the review window.

Date precision. The publisher lists a month, not an exact publication day. This report highlights the journal record.

The November 13 arXiv check matched all 676 new announcements in that batch across the eight tracked categories. The catalog indexed 4,909 distinct works over the review window, including historical records. Conference acceptance coverage on OpenReview and historical arXiv coverage remain partial.

Publication updates · Indexed works · Source health