The CoVar Zeitgeist: November, 2025

CoVar is pleased to present the CoVar Zeitgeist - our monthly overview of cutting edge-AI/ML. The November issue covers research from October 2025. Featuring:

  • A paper from Meta introducing a novel reinforcement learning framework for language models which maintains positive performance while minimizing hallucinations by encouraging the model to abstain when uncertain.

  • A novel random utility model - a correlated probit model - which, when trained on three-tuples of data, can correctly model correlations in user preferences. This has potential to improve the reward modelling in Reinforcement Learning from Human Feedback.

  • A method using multi-agent influence diagrams to target and intervene upon a single agent to affect the desired change in a multi-agent reinforcement learning framework.

  • A novel chip architecture, the thermodynamic sampling unit (TSU) which promises to greatly reduce energy expenditure by directly modelling probability distributions on hardware using pbits.

  • A novel information-theoretic approach for detecting novel out-of-distribution inputs in image data.

  • A paper which (1) establishes an equivalence between different types of AI agents and machines form the Chomsky hierarchy and (2) uses this equivalence to better characterize AI agents.

Check out the CoVar website!

LLMs

LoRA Without Regret

Thoroughly investigates the use of LoRA for post-training, and develops a training recipe which guarantees the effectiveness of LoRA. The recipe includes applying LoRA to all network layers and careful hyperparameter tuning.

TruthRL: Incentivizing Truthful LLMs via Reinforcement Learning

Designs a novel training method for LLMs which separately rewards correct answers, hallucinations, and abstentions in order to encourage abstaining over hallucinating. Reduces hallucinations while improving truthfulness.

Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models

Develops a method to automatically optimize LLM performance by iteratively refining the prompt given to the model without supervision. Allows small models to outperform large models with little cost.

How Do LLMs Use Their Depth?

Analyzes how inference develops across layers of an LLM. High frequency tokens appear in early layers and then refined with contextual information as layer depth increases.

Language Models are Injective and Hence Invertible

Proves that decoder-only transformers are almost-surely injective: unique inputs have unique representations. Leverages this to construct an algorithm that can recover the exact input prompt given hidden representations.

Novel Architectures

The Dragon Hatchling: the Missing Link between the Transformer and Models of the Brain

Introduces a novel language model architecture which employs a network of locally interacting neuron particles inspired by the biological structure of the human brain. This results in an interpretable LLM (i.e., sparse and positive activation vectors) which matches the performance of transformer-based architectures.

Platonic Transformers: a Solid Choice for Equivariance

Introduces a novel attention structure, drawing inspiration from Platonic solid symmetry groups, which is formally equivalent to a dynamic group convolution. Achieves SOTA performance.

DeepSeek-OCR: Contexts Optical Compression

Proposes a novel architecture that achieves computational efficiency increases in LLMs by converting natural language tokens to vision tokens and processing the latter.

Thermodynamic Computing: From Zero to One

Proposes a novel hardware architecture, the thermodynamic sampling unit (TSU), which directly samples probability distributions using pbits rather than using matrices of weights. This enables implementations which consume orders of magnitude less energy than current approaches.

Object Detection

Universal Beta Splatting

Proposes Universal Beta Splatting, a generalization of Gaussian Splatting which uses N-dimensional Beta kernels instead of Gaussian kernels. Achieves real time rendering and decomposition of scene objects into interpretable classes without supervision.

Bayesian Topological Convolutional Neural Nets

DEVCOM ARL proposes a novel Bayesian topological convolutional neural network architecture which can excel at computer vision tasks even when datasets are sparse and noisy.

A Multi-dimensional Semantic Surprise Framework Based on Low-Entropy Semantic Manifolds for Fine-Grained Out-of-Distribution Detection

Proposes an information-theoretic approach for out-of-distribution (OOD) identification which expands the OOD task beyond binary classification by evaluating semantic surprise using a semantic manifold. Reduces false positives rates by up to 60 percent.

Neighborhood Feature Pooling for Remote Sensing Image Classification

Develops Neighborhood Feature Pooling, a novel method for remote sensing imagery, which models local similarity relationships in feature space and enhances texture-aware classification.

Testing & Evaluation

A Geometric Approach to Optimal Experimental Design

Proposes mutual transport dependence, a novel statistical measure of dependence which is more flexible than mutual information and allows grounding in downstream applications. Leverages this measure to inform experimental design.

Benchmarking World-Model Learning

Creates a test environment for AI agents that learn world models which allows the agents to first learn the world model in an unsupervised phase before being tested on a related challenge. Models are assessed solely on results, not on internal state activations.

Stress-Testing Model Specs Reveals Character Differences among Language Models

Evaluates how frontier models handle value tradeoffs by presenting them scenarios where they must implicitly choose between different values and assessing their response.

Autonomy

Agent Learning via Early Experience

Proposes a novel paradigm for training agents via reinforcement learning, Early Experience, in domains that lack verifiable rewards, require long time horizons, or are data limited. Early Experience uses the future world states created by the agent’s own early actions to construct a reward signal.

Are Agents Just Automata? On the Formal Equivalence Between Agentic AI and the Chomsky Hierarchy

Creates an equivalence between autonomous AI agents and finite automata from the Chomsky Hierarchy. Using this equivalence, it proposes methods for characterizing AI agents.

Semantic Communication with World Models

Aims to solve the transmission problem between multiple agents by having each agent employ two modules. The first is a world foundation model which predicts how the world has developed since the transmission of the last frame, the second is a module which predicts accumulated WFM error and triggers a new frame transmission when necessary.

Task Completion Agents are not Ideal Collaborators

Shows that current agents, which are aimed at one-shot task completion, underperform in scenarios which require collaboration. Collaborative agents must develop the capability to function in a multi-turn setting, sustain user engagement, and scaffold user understanding.

The Oversight Game: Learning to Cooperatively Balance an AI Agent’s Safety and Autonomy

Proposes a minimal method to train AI agents to be aligned: have a human and the agent play an Oversight game where the human chooses whether to trust or oversee the agent, while the agent chooses to act or defer. Agents trained in this way learn when to act autonomously and when to defer.

Reinforcement Learning

h1: Bootstrapping LLMs to Reason over Long Horizons via Reinforcement Learning

Proposes novel reinforcement learning method to bootstrap a long-horizon training dataset for LLMs by combining disparate, short-horizon, tasks.

The Art of Scaling Reinforcement Learning Compute for LLMs

Devotes more than 400,000 GPU-hours to analyze scaling behavior for reinforcement learning LLMs. Develops methods to predict future gains from additional RL hours after an initial training period.

A Principle of Targeted Intervention for Multi-Agent Reinforcement Learning

Leverages Multi-Agent Influence Diagrams (MAIDs) and causal inference techniques to develop targeted interventions which function on a single agent in order to effect desired behavior.

On-Policy Distillation

A well-written explanation of what on-policy distillation is and how to best use it to have teacher models train student models.

Statistics

BALLAST: Bayesian Active Learning with Look-ahead Amendment for Sea-drifter Trajectories under Spatio-Temporal Vector Fields

Uses a spatio-temporal Gaussian Process surrogate model to improve inference for time-dependent vector fields, with an eye towards applications in predicting the trajectory of drifting seaborne objects.

Who Said Neural Networks Aren’t Linear?

Constructs vector spaces from a neural network’s domain to its range between which the neural network functions as a linear operator. This unlocks many techniques for use on neural networks, such as singular value decomposition.

Conformal Inference for Open-Set and Imbalanced Classification

Constructs a new family of conformal p-values for conformal prediction which are better behaved in the presence of novel test time labels.

Learning Correlated Reward Models: Statistical Barriers and Opportunities

The Independence of Irrelevant Alternatives (IIA) assumption, often made in random utility models (RUM) & reinforcement learning from human feedback (RLHF), collapses all human preferences to a universal utility function. This paper investigates using a correlated probit model to model RUM instead and finds that, with best-of-three preference data, the IIA assumption can be avoided.

Block Coordinate Descent for Neural Networks Provably Finds Global Minima

Proves that a block coordinate descent algorithm, when applied to a neural network, will reliably find the global minima for strictly monotonic loss functions. A second result shows a modified BCD algorithm achieves the same for ReLU.

Applications

The Universal Landscape of Human Reasoning

Creates a method, Information Flow Track (IF Track) to model the flow of human reasoning by using LLM encoders to capture entropy for uncertainty and cognitive effort along human reasoning tracks.

CoVar Seminar

Deep Reinforcement Learning in a Handful of Trials using Probabilistic Dynamics Models

Model-based RL approach which learns an ensemble of probabilistic dynamics models, which it leverages via planning methods to generate action sequences.

Partial least square regression (PLS regression)

A paper explaining the methods behind Partial Least Square Regression.

Partial least squares discriminant analysis:taking the magic away

A paper explaining the methods behind Partial Least Square Regression.

DepTR-MOT: Unveiling the Potential of Depth-Informed Trajectory Refinement for Multi-Object Tracking

Develops a method to improve multi-object tracking by incorporating depth estimates into the pipeline. Depth estimates are formed first at a frame level by using foundation models, and then distilled into a general depth estimate. Final inference does not rely on foundational models and has low latency which makes it useful for robotics applications.

Super-Resolution with Structured Motion

Develops and experimentally demonstrates super-resolution techniques for motion-blurred images of targets with sparse gradients via classical deconvolution with a motion PSF. Intensity values in motion-blurred images continuously sample over their motion path, so each pixel contains data that can’t be resolved by the same detector in a static image. By sampling the lower-resolution image in blocks to match the high-resolution size, and having a precise motion PSF, deblurring and super-resolution can be combined, and iteratively solved with Total Variation (TV) regularized least squares.