DeepSeek V4.1 Flash has claimed the top spot on Artificial Analysis's new evaluation index, and Geoffrey Hinton, known as the "Godfather of AI," has voiced support for a pause in AI development. On the research side, reports include a training method that improves generalization in robot foundation models, a technique that roughly halves the cost of agent skill optimization, and a medical AI that detects colorectal cancer from routine CT scans. Here is a roundup of the news that caught our attention today.

DeepSeek V4.1 Flash Beats Astra on AA's New Evaluation Index

Last week, the model evaluation site Artificial Analysis released its "Intelligence Index v4.3." It includes a new private evaluation that replaces the previous τ³ benchmark, and the developments around this new evaluation have become a hot topic in Reddit's LocalLLaMA community.

According to the post, GPT-6 Astra racked up a huge number of points on the new evaluation to sit alongside Fable, but DeepSeek V4.1 Flash surpassed it to take first place. The poster also pointed out that "the index was changed twice in three days, and right after it was adjusted so that Astra wouldn't look worse than Fable, a different model quietly took the number one spot."

Why it matters: the current state of affairs, where leaderboard rankings shift with every index revision, is a stark illustration of how difficult model comparison has become. We covered DeepSeek v4.1-Flash on 9/12, but with evaluation-side criteria changing and model-side evolution advancing at the same time, the very question of "what baseline should we use to judge which model is better" has become fluid.

Hinton Welcomes a Pause in AI Development — The Future of the Slowdown Debate

Geoffrey Hinton, known as the "Godfather of AI," indicated in an interview with Australia's ABC that he would welcome a pause in AI development.

Looking back, Trump dismissed calls for AI companies to slow down as a "negative force," while on the industry side, voices sympathetic to slowing development — starting with Amodei — have been spreading. Hinton has spent years warning about AI risks, and this latest remark is seen as giving the debate a push from a safety-focused standpoint.

Around the same time, an advocacy site called "Right to Intelligence" was shared in the local LLM community. It expresses concern that open-source models could be caught in the crossfire of AI safety regulation and calls for protecting the right to run AI locally. Whether we can achieve both improved safety and the freedom to run models on our own machines looks set to become a focal point of the regulatory debate ahead.

Cutting Off the "Vision Shortcut" in Robot Foundation Models — Latent Interface Training

On Hugging Face Daily Papers, a paper tackling the generalization problem of robot foundation models, "Breaking the Vision-Action Shortcut," has been drawing attention.

Robot foundation models are known to perform well within their training distribution but tend to degrade under visual distribution shifts (such as changes in camera position or lighting). One of the causes is the "vision-action shortcut" — a phenomenon in which action generation comes to depend on task-irrelevant visual cues that happened to correlate within the training data.

The proposed Latent Interface Training (LIT) consists of two stages. In the first stage, action generation is learned without images, conditioned on language, robot state, and the SE(3) end-effector pose at the end of demonstrations. In the second stage, a latent interface that aggregates visual and semantic representations is introduced and trained in a supervised manner to reconstruct that pose. Visual information is supplied to the action expert only through this interface.

Validated on four architectures — Pi0.5, MolmoAct2, FAST-WAM, and ImageWAM — the method improved success rates by 3.87 to 10.70 points on LIBERO-Plus under distribution shift. In real-world evaluations, improvements of 13.30 to 16.70 points were shown under unseen camera configurations, lighting changes, and distractors. Another notable feature is that the method does not depend on any specific architecture.

Smarter Agent Skill Optimization with Bandits — COBRA-Skills

Efforts to have LLM agents acquire their own "skills" and self-improve are spreading, but the optimization process of repeatedly trying candidate skills incurs substantial compute costs.

"COBRA-Skills" treats this skill optimization as a sequential budget allocation problem. A neural predictor estimates each skill's usefulness, and LinearUCB balances exploration and exploitation to select "which skill should be evaluated next." Combined with evolutionary operators such as re-generation, rollout mutation, and crossover, it iteratively improves the skill population.

In evaluations across 6 agent benchmarks × 3 target models, it achieved the best average performance among the compared methods while cutting optimization costs by approximately 55–58% compared to the conventional method SkillOpt. It requires only 50 optimization examples per benchmark. Interestingly, the main driver of the cost reduction was not fewer executions of the target agent, but the elimination of repeated LLM calls for trajectory analysis and skill rewriting. Consistent effects were also reported across both the Codex and Claude Code harnesses.

"Opportunistic AI" That Detects Colorectal Cancer from Routine CT Scans

In the medical field, demonstrations of "Opportunistic AI" — leveraging existing images for purposes other than the original reason for the examination — continue to accumulate. According to Radiology Business, an AI has been developed that detects colorectal cancer from non-contrast CT images captured during routine examinations.

Its significance lies in increasing opportunities for cancer detection from already-captured images without additional radiation exposure or cost. For patients with limited access to screening, it could lead to incidental cancer discovery within everyday clinical practice.

"Correcting" Noisy Preference Data Instead of Discarding It — PLC-DPO

Preference datasets used for training preference optimization (DPO) tend to contain flipped labels and ambiguous pairs. Simply excluding suspicious pairs is a common approach, but "PLC-DPO" explores a direction of correcting them rather than discarding them.

It assumes a latent state for each pair — correct, flipped, or tie — and estimates posterior weights from the margin between calibrated policy and reference models. In accordance with these weights, it composes three types of losses: reinforcement, inversion, and tie regularization. Because it reuses the log probabilities that DPO already computes, no auxiliary models or additional supervision are required.

In evaluations across 57 combinations of datasets, models, and benchmarks, it achieved an average win rate of 60.5% over DPO. With preference data quality often being the bottleneck, it stands out as an improvement with low implementation cost.

Wrapping Up

The debate over evaluation index revisions and the slowdown controversy show that the environment surrounding AI itself is in flux. At the same time, research tackling the steady, unglamorous problems — generalization, cost efficiency, data quality — continues to accumulate. Not losing sight of the balance between these two wheels looks set to become increasingly important.