AI Research
Weekly Roundup

Latent space models are widely used for analyzing high-dimensional discrete data matrices, such as patient-feature matrices in electronic health records (EHRs), by capturing complex dependence structures through low-dimensional embeddings. However, estimation becomes challenging in the imbalanced regime, where one matrix dimension is much larger than the other. In EHR applications, cohort sizes are often limited by disease prevalence or data availability, whereas the feature space remains extremely large due to the breadth of medical coding system. Motivated by the increasing availability of external semantic embeddings, such as pre-trained embeddings of clinical concepts in EHRs, we propose a knowledge-embedded latent projection model that leverages semantic side information to regularize representation learning. Specifically, we model column embeddings as smooth functions of semantic embeddings via a mapping in a reproducing kernel Hilbert space. We develop a computationally efficient two-step estimation procedure that combines semantically guided subspace construction via kernel principal component analysis with scalable projected gradient descent. We establish estimation error bounds that characterize the trade-off between statistical error and approximation error induced by the kernel projection. Furthermore, we provide local convergence guarantees for our non-convex optimization procedure. Extensive simulation studies and a real-world EHR application demonstrate the effectiveness of the proposed method.

• /The paper addresses the challenges of analyzing high-dimensional discrete data matrices, particularly in EHRs, where cohort sizes are often limited.
• /Imbalanced data regimes are a significant concern, with one dimension (e.g., patients) being much smaller than the other (e.g., features).
• /The proposed model integrates external semantic embeddings to regularize representation learning, enhancing the quality of data analysis.
• /Column embeddings are modeled as smooth functions of semantic embeddings using a mapping in a reproducing kernel Hilbert space.
• /A two-step estimation procedure is developed, combining semantically guided subspace construction via kernel principal component analysis with scalable projected gradient descent.
• /Estimation error bounds are established, characterizing the trade-off between statistical error and approximation error from kernel projection.
• /Local convergence guarantees for the non-convex optimization procedure are provided, ensuring reliable estimation in practical applications.
• /Extensive simulation studies validate the effectiveness of the proposed method in capturing complex dependence structures in EHR data.
• /The research contributes to the field of healthcare data analysis by offering a robust framework for leveraging semantic information in high-dimensional data.

LLM-based agents are increasingly being deployed in contexts requiring complex authorization policies: customer service protocols, approval workflows, data access restrictions, and regulatory compliance. Embedding these policies in prompts provides no enforcement guarantees. We present PCAS, a Policy Compiler for Agentic Systems that provides deterministic policy enforcement. Enforcing such policies requires tracking information flow across agents, which linear message histories cannot capture. Instead, PCAS models the agentic system state as a dependency graph capturing causal relationships among events such as tool calls, tool results, and messages. Policies are expressed in a Datalog-derived language, as declarative rules that account for transitive information flow and cross-agent provenance. A reference monitor intercepts all actions and blocks violations before execution, providing deterministic enforcement independent of model reasoning. PCAS takes an existing agent implementation and a policy specification, and compiles them into an instrumented system that is policy-compliant by construction, with no security-specific restructuring required. We evaluate PCAS on three case studies: information flow policies for prompt injection defense, approval workflows in a multi-agent pharmacovigilance system, and organizational policies for customer service. On customer service tasks, PCAS improves policy compliance from 48% to 93% across frontier models, with zero policy violations in instrumented runs.

• /PCAS (Policy Compiler for Agentic Systems) provides deterministic policy enforcement for LLM-based agents.
• /Traditional prompt-based policy embedding lacks enforcement guarantees, leading to compliance risks.
• /PCAS models agentic system states as dependency graphs to capture causal relationships among events.
• /Policies are expressed in a Datalog-derived language, allowing for declarative rules that consider transitive information flow.
• /A reference monitor in PCAS intercepts actions to block policy violations before execution.
• /PCAS compiles existing agent implementations with policy specifications into compliant systems without major restructuring.
• /The framework was evaluated through three case studies, demonstrating its versatility across different contexts.
• /In customer service applications, PCAS improved policy compliance from 48% to 93% with zero policy violations.
• /The research underscores the importance of robust policy enforcement in the deployment of LLM-based agents.

Visual loco-manipulation of arbitrary objects in the wild with humanoid robots requires accurate end-effector (EE) control and a generalizable understanding of the scene via visual inputs (e.g., RGB-D images). Existing approaches are based on real-world imitation learning and exhibit limited generalization due to the difficulty in collecting large-scale training datasets. This paper presents a new paradigm, HERO, for object loco-manipulation with humanoid robots that combines the strong generalization and open-vocabulary understanding of large vision models with strong control performance from simulated training. We achieve this by designing an accurate residual-aware EE tracking policy. This EE tracking policy combines classical robotics with machine learning. It uses a) inverse kinematics to convert residual end-effector targets into reference trajectories, b) a learned neural forward model for accurate forward kinematics, c) goal adjustment, and d) replanning. Together, these innovations help us cut down the end-effector tracking error by 3.2x. We use this accurate end-effector tracker to build a modular system for loco-manipulation, where we use open-vocabulary large vision models for strong visual generalization. Our system is able to operate in diverse real-world environments, from offices to coffee shops, where the robot is able to reliably manipulate various everyday objects (e.g., mugs, apples, toys) on surfaces ranging from 43cm to 92cm in height. Systematic modular and end-to-end tests in simulation and the real world demonstrate the effectiveness of our proposed design. We believe the advances in this paper can open up new ways of training humanoid robots to interact with daily objects.

• /Introduction of HERO, a new paradigm for humanoid robot loco-manipulation that enhances generalization and control performance.
• /Combines large vision models for open-vocabulary understanding with simulated training for robust end-effector control.
• /Development of a residual-aware EE tracking policy that integrates classical robotics techniques with machine learning.
• /Utilizes inverse kinematics to convert residual targets into reference trajectories, improving tracking accuracy.
• /Incorporates a learned neural forward model for precise forward kinematics, enhancing movement fidelity.
• /Features goal adjustment and replanning mechanisms to adapt to dynamic environments and tasks.
• /Achieves a 3.2x reduction in end-effector tracking error compared to previous methods.
• /Demonstrates effective manipulation of various objects in diverse real-world environments, including heights ranging from 43cm to 92cm.
• /Systematic testing in both simulation and real-world scenarios confirms the effectiveness and versatility of the HERO system.

Fast weight architectures offer a promising alternative to attention-based transformers for long-context modeling by maintaining constant memory overhead regardless of context length. However, their potential is limited by the next-token prediction (NTP) training paradigm. NTP optimizes single-token predictions and ignores semantic coherence across multiple tokens following a prefix. Consequently, fast weight models, which dynamically update their parameters to store contextual information, learn suboptimal representations that fail to capture long-range dependencies. We introduce REFINE (Reinforced Fast weIghts with Next sEquence prediction), a reinforcement learning framework that trains fast weight models under the next-sequence prediction (NSP) objective. REFINE selects informative token positions based on prediction entropy, generates multi-token rollouts, assigns self-supervised sequence-level rewards, and optimizes the model with group relative policy optimization (GRPO). REFINE is applicable throughout the training lifecycle of pre-trained language models: mid-training, post-training, and test-time training. Our experiments on LaCT-760M and DeltaNet-1.3B demonstrate that REFINE consistently outperforms supervised fine-tuning with NTP across needle-in-a-haystack retrieval, long-context question answering, and diverse tasks in LongBench. REFINE provides an effective and versatile framework for improving long-context modeling in fast weight architectures.

• /REFINE introduces a reinforcement learning approach to improve fast weight architectures for long-context modeling.
• /The framework shifts the training paradigm from next-token prediction (NTP) to next-sequence prediction (NSP), addressing limitations in capturing long-range dependencies.
• /Fast weight architectures maintain constant memory overhead, making them advantageous for processing long contexts compared to attention-based transformers.
• /REFINE selects informative token positions based on prediction entropy, allowing for the generation of multi-token rollouts.
• /The model assigns self-supervised sequence-level rewards to guide the learning process more effectively.
• /Group relative policy optimization (GRPO) is employed for optimizing the model, enhancing its performance across various tasks.
• /Experimental results demonstrate that REFINE outperforms traditional supervised fine-tuning methods based on NTP in multiple benchmarks.
• /The framework is applicable throughout the training lifecycle, including mid-training, post-training, and test-time training.
• /REFINE shows significant improvements in tasks such as needle-in-a-haystack retrieval and long-context question answering, indicating its versatility and effectiveness.

Large language models (LLMs) perform strongly on biological benchmarks, raising concerns that they may help novice actors acquire dual-use laboratory skills. Yet, whether this translates to improved human performance in the physical laboratory remains unclear. To address this, we conducted a pre-registered, investigator-blinded, randomized controlled trial (June-August 2025; n = 153) evaluating whether LLMs improve novice performance in tasks that collectively model a viral reverse genetics workflow. We observed no significant difference in the primary endpoint of workflow completion (5.2% LLM vs. 6.6% Internet; P = 0.759), nor in the success rate of individual tasks. However, the LLM arm had numerically higher success rates in four of the five tasks, most notably for the cell culture task (68.8% LLM vs. 55.3% Internet; P = 0.059). Post-hoc Bayesian modeling of pooled data estimates an approximate 1.4-fold increase (95% CrI 0.74-2.62) in success for a "typical" reverse genetics task under LLM assistance. Ordinal regression modelling suggests that participants in the LLM arm were more likely to progress through intermediate steps across all tasks (posterior probability of a positive effect: 81%-96%). Overall, mid-2025 LLMs did not substantially increase novice completion of complex laboratory procedures but were associated with a modest performance benefit. These results reveal a gap between in silico benchmarks and real-world utility, underscoring the need for physical-world validation of AI biosecurity assessments as model capabilities and user proficiency evolve.

• /The study was a pre-registered, investigator-blinded, randomized controlled trial involving 153 participants.
• /The primary endpoint assessed was the completion rate of a viral reverse genetics workflow, with no significant difference found between LLM assistance and traditional internet resources.
• /The overall workflow completion rates were 5.2% for the LLM group and 6.6% for the internet group (P = 0.759).
• /Numerically higher success rates were observed in the LLM group for four out of five tasks, particularly in the cell culture task (68.8% LLM vs. 55.3% Internet; P = 0.059).
• /Post-hoc Bayesian modeling estimated a 1.4-fold increase in success for typical reverse genetics tasks with LLM assistance, although with a wide credible interval (95% CrI 0.74-2.62).
• /Ordinal regression modeling indicated that LLM participants were more likely to progress through intermediate steps across all tasks, with a posterior probability of a positive effect between 81% and 96%.
• /The results suggest a modest performance benefit from LLM assistance, despite not achieving significant improvements in overall task completion.
• /The study underscores the gap between in silico performance benchmarks and practical laboratory outcomes.
• /It emphasizes the need for ongoing validation of AI tools in real-world settings, particularly in the context of biosecurity and laboratory skills acquisition.

Vision-language models (VLMs) aim to reason by jointly leveraging visual and textual modalities. While allocating additional inference-time computation has proven effective for large language models (LLMs), achieving similar scaling in VLMs remains challenging. A key obstacle is that visual inputs are typically provided only once at the start of generation, while textual reasoning (e.g., early visual summaries) is generated autoregressively, causing reasoning to become increasingly text-dominated and allowing early visual grounding errors to accumulate. Moreover, vanilla guidance for visual grounding during inference is often coarse and noisy, making it difficult to steer reasoning over long texts. To address these challenges, we propose \emph{Saliency-Aware Principle} (SAP) selection. SAP operates on high-level reasoning principles rather than token-level trajectories, which enable stable control over discrete generation under noisy feedback while allowing later reasoning steps to re-consult visual evidence when renewed grounding is required. In addition, SAP supports multi-route inference, enabling parallel exploration of diverse reasoning behaviors. SAP is model-agnostic and data-free, requiring no additional training. Empirical results show that SAP achieves competitive performance, especially in reducing object hallucination, under comparable token-generation budgets while yielding more stable reasoning and lower response latency than CoT-style long sequential reasoning.

• /Vision-language models (VLMs) struggle with visual grounding due to the static nature of visual inputs at the start of generation.
• /Textual reasoning in VLMs often becomes dominant, leading to compounded errors from initial visual grounding mistakes.
• /Existing visual grounding guidance methods are typically coarse and noisy, complicating long-text reasoning.
• /The Saliency-Aware Principle (SAP) selection method is proposed to enhance reasoning by focusing on high-level principles rather than token-level details.
• /SAP allows for stable control over discrete generation processes, even under noisy feedback conditions.
• /The method enables later reasoning steps to re-consult visual evidence, improving accuracy when renewed grounding is necessary.
• /SAP supports multi-route inference, facilitating parallel exploration of diverse reasoning behaviors and enhancing output richness.
• /The approach is model-agnostic and does not require additional training, making it adaptable to various VLM architectures.
• /Empirical results show that SAP reduces object hallucination and provides more stable reasoning with lower response latency compared to traditional CoT methods.

LLMs are increasingly being used for complex problems which are not necessarily resolved in a single response, but require interacting with an environment to acquire information. In these scenarios, LLMs must reason about inherent cost-uncertainty tradeoffs in when to stop exploring and commit to an answer. For instance, on a programming task, an LLM should test a generated code snippet if it is uncertain about the correctness of that code; the cost of writing a test is nonzero, but typically lower than the cost of making a mistake. In this work, we show that we can induce LLMs to explicitly reason about balancing these cost-uncertainty tradeoffs, then perform more optimal environment exploration. We formalize multiple tasks, including information retrieval and coding, as sequential decision-making problems under uncertainty. Each problem has latent environment state that can be reasoned about via a prior which is passed to the LLM agent. We introduce a framework called Calibrate-Then-Act (CTA), where we feed the LLM this additional context to enable it to act more optimally. This improvement is preserved even under RL training of both the baseline and CTA. Our results on information-seeking QA and on a simplified coding task show that making cost-benefit tradeoffs explicit with CTA can help agents discover more optimal decision-making strategies.

• /LLMs are increasingly utilized for complex tasks requiring interaction with environments, necessitating advanced reasoning capabilities.
• /The research identifies a critical need for LLMs to balance cost-uncertainty tradeoffs when deciding to explore or commit to an answer.
• /An example scenario involves programming tasks where testing code snippets incurs costs that must be weighed against the risks of errors.
• /The study formalizes various tasks, including information retrieval and coding, as sequential decision-making problems under uncertainty.
• /The Calibrate-Then-Act (CTA) framework is introduced, which provides LLMs with additional contextual information to enhance decision-making.
• /The methodology allows LLMs to reason about latent environmental states, improving their ability to make informed choices.
• /Empirical results indicate that the CTA framework leads to more optimal exploration strategies in LLMs.
• /The improvements in decision-making persist even under reinforcement learning training, showcasing the robustness of the CTA approach.
• /This research has significant implications for the development of AI systems that require complex reasoning and optimal problem-solving capabilities.

Interpretability research on large language models (LLMs) has yielded important insights into model behaviour, yet recurring pitfalls persist: findings that do not generalise, and causal interpretations that outrun the evidence. Our position is that causal inference specifies what constitutes a valid mapping from model activations to invariant high-level structures, the data or assumptions needed to achieve it, and the inferences it can support. Specifically, Pearl's causal hierarchy clarifies what an interpretability study can justify. Observations establish associations between model behaviour and internal components. Interventions (e.g., ablations or activation patching) support claims how these edits affect a behavioural metric (\eg, average change in token probabilities) over a set of prompts. However, counterfactual claims -- i.e., asking what the model output would have been for the same prompt under an unobserved intervention -- remain largely unverifiable without controlled supervision. We show how causal representation learning (CRL) operationalises this hierarchy, specifying which variables are recoverable from activations and under what assumptions. Together, these motivate a diagnostic framework that helps practitioners select methods and evaluations matching claims to evidence such that findings generalise.

• /The paper critiques the current state of interpretability research in large language models, identifying key pitfalls such as non-generalizable findings and unsupported causal claims.
• /It emphasizes the role of causal inference in establishing valid mappings from model activations to high-level structures, which are invariant across different contexts.
• /The authors utilize Judea Pearl's causal hierarchy to clarify the limitations of interpretability studies, distinguishing between observations and interventions.
• /Interventions like ablations and activation patching are discussed as methods to demonstrate how changes affect behavioral metrics, such as token probabilities.
• /A significant focus is placed on the challenge of making counterfactual claims without controlled supervision, highlighting a gap in current methodologies.
• /The concept of causal representation learning (CRL) is introduced as a means to operationalize the causal hierarchy, specifying recoverable variables from activations.
• /The paper proposes a diagnostic framework to assist practitioners in aligning methods and evaluations with evidence, enhancing the reliability of findings.
• /The research aims to improve the generalizability of interpretability studies, advocating for a structured approach to causal inference in AI.
• /Overall, the paper contributes to the broader understanding of model interpretability, pushing for a more evidence-based methodology in the analysis of large language models.

Single-cell RNA sequencing (scRNA-seq) data exhibit strong and reproducible statistical structure. This has motivated the development of large-scale foundation models, such as TranscriptFormer, that use transformer-based architectures to learn a generative model for gene expression by embedding genes into a latent vector space. These embeddings have been used to obtain state-of-the-art (SOTA) performance on downstream tasks such as cell-type classification, disease-state prediction, and cross-species learning. Here, we ask whether similar performance can be achieved without utilizing computationally intensive deep learning-based representations. Using simple, interpretable pipelines that rely on careful normalization and linear methods, we obtain SOTA or near SOTA performance across multiple benchmarks commonly used to evaluate single-cell foundation models, including outperforming foundation models on out-of-distribution tasks involving novel cell types and organisms absent from the training data. Our findings highlight the need for rigorous benchmarking and suggest that the biology of cell identity can be captured by simple linear representations of single cell gene expression data.

• /The study focuses on single-cell RNA sequencing (scRNA-seq) data, which exhibits strong statistical structure.
• /Large-scale foundation models like TranscriptFormer utilize transformer architectures to create generative models for gene expression.
• /These models have achieved state-of-the-art performance in tasks such as cell-type classification and disease-state prediction.
• /The research investigates whether similar performance can be attained using simpler, interpretable methods instead of deep learning.
• /The authors employed careful normalization and linear methods to analyze scRNA-seq data.
• /Results indicate that these simpler methods achieved state-of-the-art or near state-of-the-art performance across multiple benchmarks.
• /The study highlights the ability of these methods to outperform foundation models on out-of-distribution tasks involving novel cell types and organisms.
• /The findings suggest a need for rigorous benchmarking in the field of single-cell analysis.
• /The research proposes that the biological complexities of cell identity can be captured using linear representations, challenging the necessity of deep learning approaches.

Multiple hypothesis testing with false discovery rate (FDR) control is a fundamental problem in statistical inference, with broad applications in genomics, drug screening, and outlier detection. In many such settings, researchers may have access not only to real experimental observations but also to auxiliary or synthetic data -- from past, related experiments or generated by generative models -- that can provide additional evidence about the hypotheses of interest. We introduce SynthBH, a synthetic-powered multiple testing procedure that safely leverages such synthetic data. We prove that SynthBH guarantees finite-sample, distribution-free FDR control under a mild PRDS-type positive dependence condition, without requiring the pooled-data p-values to be valid under the null. The proposed method adapts to the (unknown) quality of the synthetic data: it enhances the sample efficiency and may boost the power when synthetic data are of high quality, while controlling the FDR at a user-specified level regardless of their quality. We demonstrate the empirical performance of SynthBH on tabular outlier detection benchmarks and on genomic analyses of drug-cancer sensitivity associations, and further study its properties through controlled experiments on simulated data.

• /SynthBH is a synthetic-powered multiple testing procedure designed to control the false discovery rate (FDR) in statistical inference.
• /The method leverages auxiliary or synthetic data from past experiments or generative models to enhance hypothesis testing.
• /SynthBH guarantees finite-sample, distribution-free FDR control under a mild PRDS-type positive dependence condition.
• /The method does not require pooled-data p-values to be valid under the null hypothesis, expanding its applicability.
• /SynthBH adapts to the quality of synthetic data, improving sample efficiency and statistical power when data quality is high.
• /It maintains FDR control at a user-specified level, regardless of the quality of synthetic data used.
• /Empirical performance is demonstrated on tabular outlier detection benchmarks and genomic analyses of drug-cancer sensitivity.
• /Controlled experiments on simulated data further investigate the properties and effectiveness of SynthBH.
• /The introduction of SynthBH represents a significant advancement in multiple hypothesis testing, particularly in high-dimensional data analysis.