AI papers — 2026-10-17

The Morning Preprint — this day as one edition: the day's papers, the five things worth knowing, and the episodes behind them.

Today's focus is on how we can make the agents we deploy actually adapt to changing business rules instead of just being static systems. Current methods for building and evaluating these agents do not offer a way to fix their behavior after they are live without rewriting the core code, which is too slow and costly.

We are looking at Agent Gym, a framework designed to put any existing LLM agent into a continuous evaluation-and-evolution loop. This system uses six capabilities—Act, Evaluate, Investigate, Correct, Learn, and Observe—to allow subject matter experts to discover and validate new correction rules through natural language interaction.

This framework builds on the idea of a hybrid deterministic-LLM correction engine featuring twenty-one condition operators and three tiers of actions for making these corrections. This is supported by a three-layer investigation architecture for checking compliance without needing ground truth, which is then connected to the learning loop where experts refine rules.

The concept of system transparency is also addressed through the Spec-to-Note Gap, which uses an autoencoder view to understand how agents operate. This continuous refinement process complements work on improving guidance for multi-turn agents, where State-Matched Routing and Contextualized Self-Distillation helps guide the student agent when a reference might not perfectly align with its current state.

Furthermore, when we look at document selection, Rubric4Setwise proposes a training-free method that converts rubric criteria into signals for selecting document sets, achieving top performance by closing the loop from evaluation to optimization. This contrasts with other methods that struggle with inter-document interactions.

The work on linear strategic classification with endogenous improvements is particularly important because it moves beyond treating strategic responses as mere cosmetic changes in features, allowing agents to strategically alter outcomes at a cost. This means we can model how choices affect the final result, which is crucial for understanding decision-making in complex systems. The core idea of AKMPC is that under specific assumptions about linear classifiers and costs, agents' best responses simplify to a closed-form shift of the decision boundary, making the resulting improvement-aware classification problem learnable with linear sample complexity.

This concept builds upon earlier work concerning how models evolve or are structured. For instance, BrainDyn BrainDyn introduces a graph ordinary differential equation network built on a cellular sheaf to model brain activity dynamics, using learned linear transformations to mediate region-to-region communication, and this structural learning is relevant when considering how agents might adapt their internal representations over time.

Another area where structure guides performance is in forecasting. Njord Njord presents a probabilistic graph neural network for ocean forecasting that uses a deep latent variable framework to sample forecasts, demonstrating strong performance compared to deterministic models by providing uncertainty estimates from the ensemble of sampled forecasts.

Furthermore, when we consider how agents interact with large language models, Mind the Delay Mind the Delay introduces a latency-aware query router that uses a Batch Evolution Aware estimator to predict time-to-first-token, which is vital for balancing accuracy and cost when routing queries across heterogeneous LLM pools.

Finally, the work on aligning models with human intent touches upon how we can ensure robust behavior. Mind the Gap Mind the Gap shows that standard surrogate losses for aligning LLMs are flawed, proposing a margin-shifted ranking framework that leads to SA-DPO consistently outperforming DPO and SimPO in downstream generation quality.

The most significant work today involves developing a benchmark for Structure-from-Motion methods for novel view synthesis because it directly addresses the efficiency versus accuracy trade-off in creating 3D scene representations from images. This is important because as novel view synthesis techniques like NeRFs and 3DGS become more common, having reliable and fast camera pose estimation is crucial for practical applications. The paper by AKMPC tackles this by creating a benchmark focusing on SfM for novel view synthesis, showing that combining feed-forward networks for initial estimates with classical SfM refinement yields the best efficiency-effectiveness balance.

This is complemented by research into improving large language model training data, where Composition-RL proposes an offline data augmentation approach to handle the bottleneck caused by easy prompts in Reinforcement Learning with Verifiable Rewards. This method composes existing prompts before RL training, and experiments across models from 4B to 30B show that this consistently improves reasoning capability over training on the original dataset.

Another area of advancement focuses on making generative models faster, as seen in Sig-DEG, which introduces a generator for distilling pre-trained diffusion models. This technique uses partial signatures to approximate solutions to stochastic differential equations, enabling inference speedups of up to several hundred times over the teacher model while maintaining competitive sample quality.

Furthermore, work on remote sensing multimodal large language models is pushing the boundaries of scenario specialization with Filling Before Advancing. This approach organizes supervision into semantic anchoring, multi-source bridging, and evidence-grounded tuning to develop prerequisite capabilities before adapting to specific scenarios like coastal harbor understanding.

Finally, in the realm of classification loss functions, BOLT introduces a new loss function called Bayes optimal learning threshold which computes an expectation upper bound on the Bayes error rate from classifier outputs. This allows model parameters to be updated toward minimizing this bound during training, often resulting in lower held-out classification errors compared to standard cross-entropy training.

The most significant work today concerns the development of EPIC-Bench, which matters because it provides a systematic way to test how well large vision-language models understand real-world embodied situations, specifically in areas like target localization and manipulation. This benchmark gathers six thousand six hundred human-annotated examples across three stages of interaction—Target Localization, Navigation, and Manipulation—to diagnose the current limitations of these models.

This diagnostic work is supported by several related efforts. Naka-GS tackles the problem of low-light 3D Gaussian Splatting, which suffers from color distortion and geometric errors when inputs are dim. This framework introduces a method called Naka-Guided Chroma Correction to learn luminance and chromatic adjustments while preserving fine structures, leading to improved reconstruction quality in scenes tested on eight scene-disjoint RealX3D scenes.

Another piece of research addresses the physical plausibility of generated 3D scenes through PhyMix, which incorporates feedback from a Physics Evaluator into both training and inference. This framework uses implicit alignment via Scene-GRPO and explicit refinement via verifier-guided agentic repair to ensure generated layouts adhere to real-world constraints.

The broader context involves how language models handle complex reasoning and bias. Reasoning Error from Known Fact explores how LLMs make factual mistakes during long reasoning traces, proposing Step-level Self-Consistency Group Relative Policy Optimization to mitigate context-sensitive hallucinations. Simultaneously, MBBQ creates a dataset for comparing stereotypes across different languages, showing that some non-English models exhibit more bias than English ones when prompted.

Finally, there is work on improving motion understanding through FiGMo, which represents motion as timestamped skeletal poses to allow reasoning over action order and duration. This approach helps preserve physical timing even when temporal subsampling occurs in video data.

The work by EvoFlux is particularly important because it directly addresses the failure modes of small language models when they are tasked with using external tools in real-world scenarios; it shows how to fix those execution errors before they derail an agent's plan. This method works by running an evolutionary loop over structured graphs of tool calls, where execution mistakes guide targeted repairs like changing a tool or fixing a parameter, and this approach significantly boosts success rates from about three percent up to seventeen to twenty-four percent on complex tasks across various planners.

This concept of iterative repair is supported by the work in EvoFlux, which shows how inference-time search can fix plans using execution errors as feedback, and this is complemented by SimCT, which tackles the problem of restoring supervision when models use different tokenizers during on-policy distillation by comparing multi-token continuations. Furthermore, Rethinking Calibration for High-Sparsity LLM Pruning explores how to better select pruning methods by treating calibration as a capability coverage problem rather than relying on a single source, suggesting that balanced mixtures outperform individual sources in retaining diverse capabilities.

The framework for safety reasoning in underground mining also provides context for structured knowledge representation, as the work on From 3D Perception to Safety Reasoning uses scene and temporal graphs to link perception outputs with reasoning stages, showing how layered analysis can increase hazard identification accuracy. This idea of structured reasoning is also touched upon in Explanation Generation for Contradiction Reconciliation with LLMs, which focuses on generating explanations that make contradictory statements compatible rather than just choosing one statement over the other.

The most significant work from yesterday involves C-VARC, which addresses the critical need for culturally grounded resources to align large language models with Chinese values by providing a corpus of rules of thumb and derived values AKMPC. This resource is vital because it moves beyond abstract guidance to offer concrete behavioral direction, which has shown strong performance in model alignment tasks, specifically outperforming matched value-only supervision on CMoralEval for both Qwen3-8B and Llama-3.1-8B-Instruct with notable gains of 2.97 and 12.82 percentage points respectively AKMPC.

Moving to memory solutions, MegaMem tackles the challenge of giving ultra-large language models persistent memory for massive contexts like entire codebases or long interaction histories MegaMem. This system separates semantic access from generation evidence, allowing it to search hundreds of millions to one billion tokens while only feeding a bounded source evidence to the answer model MegaMem.

In terms of generative models, Equilibrium Forcing introduces a framework for video denoising that decouples training and inference by removing noise level conditioning Equilibrium Forcing. This flexibility allows inference algorithms to adapt to feedback from the sample, leading to improved video quality on challenging autoregressive benchmarks Equilibrium Forcing.

The work on making transformers logical suggests a way to correct systematic errors in generalization by initializing the additive attention mask directly with task-level interaction structure rather than using fixed projections Making Transformers Logical. This approach allows the structural prior to persist throughout training, achieving near-perfect extrapolation on Boolean reasoning tasks Making Transformers Logical.

Finally, the research into causal deep reinforcement learning proposes MOSE, which constructs a provably minimal state by exposing a single shared Q-function to history-based state constructions of multiple orders simultaneously From Causal Graphs to Markovian States. This method shows that controlled redundancy gives the model flexibility to learn its state representation end-to-end, outperforming minimal state construction alone From Causal Graphs to Markovian States.

The work on Writer-R1 is particularly important because it tackles the fundamental challenge of creating reliable specifications for open-ended writing tasks at scale. This pipeline builds a system around Inductive Rubric Generation to automatically construct the criteria needed for training models in reinforcement learning, which means the requirements for good writing are made available to the writer before they even generate text. This approach improves WritingBench scores by gaining 0.182 with the 4B backbone and 0.158 with the 8B backbone, and these gains also extend to HelloBench, suggesting that using these inductively mined rubrics as both guidance and reward specifications is a viable path forward for open-ended writing tasks.

The Rational Transconductors paper addresses a core limitation in standard Transformers regarding rigid sequential logic by introducing a dual-stream architecture that incorporates matrix-valued recurrence derived from Weighted Finite Automata. This method injects rational state information into the attention mechanism via a Deep Rational Injection scheme, which allows the framework to strictly generalize Transformer expressive power to capture all Regular Languages and 1-complete languages while maintaining parallel training efficiency. This theoretical grounding is supported by proofs showing that Random Rational Features serve as a universal basis for sequential dependencies, justifying their initialization strategy.

TASTE proposes a task-aware framework using Stein operators to detect out-of-distribution shifts in models, linking distribution shift directly to the input sensitivity of the model rather than just looking at the data itself. This method allows for geometric interpretation of distribution shift and enables localization through coordinate-wise decomposition, showing competitive performance on benchmarks like MNIST and ImageNet1K under controlled Gaussian shifts.

Can MLLMs Read the Room? A Multimodal Benchmark for Assessing Deception in Multi-Party Social Interactions introduces MIDA to rigorously test if state-of-the-art Multimodal Large Language Models possess the ability to assess deception in social settings. The findings reveal a significant performance gap, showing that powerful models like GPT4o struggle to reliably distinguish truthful statements from deceptive ones because they have difficulty grounding language in multimodal social cues. To combat this, the authors propose a Social Chain-of-Thought reasoning pipeline and a Dynamic Social Epistemic Memory module, finding that structured epistemic memory consistently improves deception assessment.

Thinking-Free Policy Initialization makes Distilled Reasoning Models More Effective and Efficient Reasoners addresses the computational burden of Reinforcement Learning with Verifiable Reward by introducing Thinking-Free Policy Initialization. This technique discards thinking content via a direct think append operation, which accelerates RL convergence and achieves higher performance ceilings, such as 89.0% accuracy on AIME24 for a 4B model, while significantly lowering token consumption.

MEMONDEMAND introduces a memory management system for large-scale enterprise data repositories that coordinates dynamic multi-level hierarchies and dual memory at every level to handle retrieval efficiently. This system outperforms the strongest published results on EnterpriseRAG-Bench across various data scales, demonstrating strong performance in fact-retrieval settings like FinanceBench and HotpotQA.

PureTD revisits Tesauro's TD-Gammon for backgammon money games by using pure self-play reinforcement learning without requiring evaluation-time search. This method shows that training models via cubeful self-play suffices to reach near state-of-the-art playing strength, outperforming engines relying on one move lookahead search.

SkillCommit proposes an online skill evolution framework that transforms individual experiences into a hierarchical library of reusable skills by abstracting behaviorally compatible strategies. This system preserves experience as instance-specific patches before abstracting them into higher-level skills, which allows learned skills to transfer across model scales and families, improving agent performance on benchmarks like RuleArena.

The most significant finding relates to AdaPCLA, which shows consistent gains in tail plausibility and downstream utility when generating electronic health records by adapting training strategies based on data distribution knowledge AdaPCLA. This matters because standard models often fail to capture the structure of rare events, which is crucial for studying diseases or symptoms in small groups. This adaptive fitting, achieved through simulated annealing training and zero-shot control, allows generative models to handle diverse clinical populations effectively.

The work on Adaptive Compute in Latent World Models suggests that choosing the right depth per step in a latent predictor can reduce rollout error significantly compared to fixed depths or schedules Does Depth Survive Composition, and Can a Router Use It?. This means that dynamically adjusting computational resources based on the required level of detail during prediction leads to better performance in complex planning tasks.

Tail-Aware Credit Calibration addresses a specific failure mode in reinforcement learning where uniform credit assignment can reinforce incorrect rare tokens Right Answer, Wrong Tokens: Tail-Aware Credit Calibration for RLVR. By using the policy itself to assess token rarity, this method softly reduces positive credit for high tail-risk tokens without eliminating learning opportunities.

UniWind provides a framework that integrates physics-informed priors with data-driven corrections for wind power forecasting UniWind: Day-Ahead Wind Power Forecasting via Physics-Informed State Routing. This approach uses a physical prior to set constraints on generation potential, which is then refined by state-aware corrections based on historical discrepancies and future weather context.

Learning at the Right Pace proposes Adaptive Data Scheduling to improve reinforcement learning accuracy by organizing samples semantically and selecting them near the policy's capability boundary Learning at the Right Pace: Adaptive Data Scheduling Improves LLM Reinforcement Learning. This scheduling method helps LLMs learn more effectively by focusing on data that pushes their current limits rather than uniform sampling.

Finally, the study on Unification and Optimization of Robust Supervised Learning provides a framework for organizing various robustness methods along explicit design axes Unification and Optimization of Robust Supervised Learning. This work helps researchers understand how different uncertainty handling techniques interact, showing that the best approach often depends on the specific task and data used.

Today's papers

The papers

Important terms

Agent Gym
A framework that puts any existing LLM agent into a continuous loop of evaluation and evolution using six capabilities like Act, Evaluate, and Learn to adapt to changing business rules.
AKMPC
A concept for linear strategic classification where agents can strategically alter outcomes at a cost. It simplifies the problem by modeling choices as shifts in the decision boundary.
EvoFlux
A method that fixes execution errors in agent plans by running an evolutionary loop over tool calls, using mistakes to guide targeted repairs and boosting success rates.
EPIC-Bench
A major benchmark for vision-language models that systematically tests their understanding of real-world embodied situations like target localization and manipulation.