We propose a novel method that leverages sparse autoencoders (SAEs) and clustering techniques to analyze the internal token representations of large language models (LLMs) and guide generations in mathematical reasoning tasks. Our approach first trains an SAE to generate sparse vector representations for training tokens, then applies k-means clustering to construct a graph where vertices represent token clusters and weighted edges capture sequential token transitions. Using this graph, we define an edge-weight based reward function to quantify adherence to established reasoning traces, thereby identifying exploitative reasoning trajectories. Additionally, we measure generation diversity from clustering to assess the extent of exploration. Our findings indicate that balancing both exploitation and exploration is crucial for achieving high accuracy in mathematical reasoning tasks. During generation, the SAE can serve as a scalable reward model to guide generations, ensuring a balanced trade-off between exploitation and exploration. This prevents extreme behaviors in either direction, ultimately fostering a higher-quality reasoning process in LLMs.

Large Language Models (LLMs) are emerging as versatile foundation models for computational chemistry, handling bidirectional tasks like reaction prediction and retrosynthesis. However, these models often lack round-trip consistency. For instance, a state-of-the-art chemical LLM may successfully caption a molecule, yet be unable to accurately reconstruct the original structure from its own generated text. This inconsistency suggests that models are learning unidirectional memorization rather than flexible mastery. Indeed, recent work has demonstrated a strong correlation between a model's round-trip consistency and its performance on the primary tasks. We therefore introduce Round-Trip Reinforcement Learning (RTRL), a novel framework that trains a model to improve its consistency by using the success of a round-trip transformation as a reward signal. We further propose an iterative variant where forward and reverse mappings alternately train each other in a self-improvement loop, a process that is highly data-efficient and notably effective with the massive amount of unlabelled data common in chemistry. Experiments demonstrate that RTRL significantly boosts performance and consistency over strong baselines across supervised, self-supervised, and synthetic data regimes. This work shows that round-trip consistency is not just a desirable property but a trainable objective, offering a new path toward more robust and reliable foundation models. Large Language Models (LLMs) are emerging as a powerful class of versatile and generalizable foundation models for computational chemistry (Fang et al.; Cao et al., 2025; Wang et al., 2025). A key advantage is their ability to provide a more flexible and intuitive interface for scientists, enabling interaction with complex data through natural language. This is powered by their unique capability to process and generate information across diverse chemical modalities, from structured SMILES strings to unstructured experimental procedures (Zhao et al., 2025; Pei et al., 2024a). By unifying these disparate tasks and data types within a single framework, LLMs promise to significantly accelerate the pace of scientific discovery. A prominent theme enabled by this new generation of models is the modeling of bidirectional chemical tasks, such as translating between textual descriptions and molecular structures or mapping between forward synthesis and retrosynthesis (Pei et al., 2024b; Edwards et al., 2024; Liu et al., 2024). More than just a capability, however, this bidirectionality allows us to examine a model's depth of understanding through the crucial property of round-trip consistency (RTC).

Domain-specific quantitative reasoning remains a major challenge for large language models (LLMs), especially in fields requiring expert knowledge and complex question answering (QA). In this work, we propose Expert Question Decomposition (EQD), an approach designed to balance the use of domain knowledge with computational efficiency. EQD is built on a two-step fine-tuning framework and guided by a reward function that measures the effectiveness of generated sub-questions in improving QA outcomes. It requires only a few thousand training examples and a single A100 GPU for fine-tuning, with inference time comparable to zero-shot prompting. Beyond its efficiency, EQD outperforms state-of-the-art domain-tuned models and advanced prompting strategies. We evaluate EQD in the financial domain, characterized by specialized knowledge and complex quantitative reasoning, across four benchmark datasets. Our method consistently improves QA performance by 0.6% to 10.5% across different LLMs. Our analysis reveals an important insight: in domain-specific QA, a single supporting question often provides greater benefit than detailed guidance steps.

Web agents promise to automate complex browser tasks, but current methods remain brittle -- relying on step-by-step UI interactions and heavy LLM reasoning that break under dynamic layouts and long horizons. Humans, by contrast, exploit website-provided functionality through high-level operations like search, filter, and sort. We introduce WALT (Web Agents that Learn Tools), a framework that reverse-engineers latent website functionality into reusable invocable tools. Rather than hypothesizing ad-hoc skills, WALT exposes robust implementations of automations already designed into websites -- spanning discovery (search, filter, sort), communication (post, comment, upvote), and content management (create, edit, delete). Tools abstract away low-level execution: instead of reasoning about how to click and type, agents simply call search(query) or create(listing). This shifts the computational burden from fragile step-by-step reasoning to reliable tool invocation. On VisualWebArena and WebArena, WALT achieves higher success with fewer steps and less LLM-dependent reasoning, establishing a robust and generalizable paradigm for browser automation.

Analysis of multi-modal content can be tricky, computationally expensive, and require a significant amount of engineering efforts. Lots of work with pre-trained models on static data is out there, yet fusing these opensource models and methods with complex data such as videos is relatively challenging. In this paper, we present a framework that enables efficiently prototyping pipelines for multi-modal content analysis. W e craft a candidate recipe for a pipeline, marrying a set of pre-trained models, to convert videos into a temporal semi-structured data format. W e translate this structure further to a frame-level indexed knowledge graph representation that is query-able and supports continual learning, enabling the dynamic incorporation of new domain-specific knowledge through an interactive medium.

Modern deployments increasingly allocate large test-time compute (thousands of tokens or many node expansions) to boost reliability. Under such budgets, standard Tree-of-Thoughts-style search exhibits two pathologies: breadth saturation (additional samples mostly produce near-duplicates, so width stops growing) and depth myopia (noisy short-horizon utilities prune branches whose payoff appears after a few more steps). We propose Lateral Tree-of-Thoughts (LToT), a drop-in controller that separates utility from logical consistency and treats low-utility but consistent candidates as assets rather than waste. The frontier is split into mainlines (high-utility candidates used for exploitation) and laterals (consistent, initially low-utility candidates that receive short, cheap probes before judgment). LToT explores laterals via Lateral Racing with Short-Circuit (LR--SC): a capped successive-halving race that spreads tiny probes across a very wide lateral set, uses width-aware thresholds with repeat-to-confirm, and immediately promotes a branch once its envelope clears the mainline bar; mainlines are kept intentionally narrow so surplus compute is invested where width is cheap. We prove a pseudolinear lateral cost $Θ(N_0 \log_η N_0)$ with logarithmically many rungs (initial lateral width $N_0$; culling factor $η>1$), in contrast to the exponential growth of uncapped mainlines. Empirical evaluations on benchmark tasks are in preparation and will be added in a future revision. In short, LToT turns large test-time budgets into principled diversity while preserving promotion discipline, mitigating saturation and myopia without inflating compute.

The aggregation of responses from multiple large language models has been widely used in practice. For example, a popular application is to improve reasoning via multi-agent LLM debate [Khan et al., 2024, Subramaniam et al., 2025, Choi et al., 2025] and LLM council [Zhao et al., 2024]. Previous works thus far have mostly employed the simple majority voting (MV) rule as a natural first instinct to aggregate different LLMs' responses into a single answer. Intuitively, MV can be viewed as a zero-order aggregation method that only depends on the observed answers and fails to account for heterogeneity and correlation among models, which are often captured by higher-order information such as LLMs' expected accuracies (first-order information) and answer correlation (second-order information). This thus raises the following natural question: is it possible to leverage such higher-order information to develop better methods for aggregating LLMs' responses?

Vision-language models (VLMs) have shown potential for robot navigation but encounter fundamental limitations: they lack persistent scene memory, offer limited spatial reasoning, and do not scale effectively with video duration for real-time application. We present VL-KnG, a Visual Scene Understanding system that tackles these challenges using spatiotemporal knowledge graph construction and computationally efficient query processing for navigation goal identification. Our approach processes video sequences in chunks utilizing modern VLMs, creates persistent knowledge graphs that maintain object identity over time, and enables explainable spatial reasoning through queryable graph structures. We also introduce WalkieKnowledge, a new benchmark with about 200 manually annotated questions across 8 diverse trajectories spanning approximately 100 minutes of video data, enabling fair comparison between structured approaches and general-purpose VLMs. Real-world deployment on a differential drive robot demonstrates practical applicability, with our method achieving 77.27% success rate and 76.92% answer accuracy, matching Gemini 2.5 Pro performance while providing explainable reasoning supported by the knowledge graph, computational efficiency for real-time deployment across different tasks, such as localization, navigation and planning. Code and dataset will be released after acceptance.

This matching requires researchers to know the exact wording that other researchers previously used, creating a challenging process that could lead to missing relevant data. Large Language Models (LLMs) could enhance data discovery by removing this requirement and allowing researchers to ask questions with natural language. However, we do not currently know if researchers would accept LLMs for data discovery. Using a human-centered artificial intelligence (HCAI) focus, we ran focus groups (N = 27) to understand researchers' perspectives towards LLMs for data discovery. Our conceptual model shows that the potential benefits are not enough for researchers to use LLMs instead of current technology. Barriers prevent researchers from fully accepting LLMs, but features around transparency could overcome them. Using our model will allow developers to incorporate features that result in an increased acceptance of LLMs for data discovery.

Data from online job postings are difficult to access and are not built in a standard or transparent manner. Data included in the standard taxonomy and occupational information database (O*NET) are updated infrequently and based on small survey samples. We adopt O*NET as a framework for building natural language processing tools that extract structured information from job postings. We publish the Job Ad Analysis Toolkit (JAAT), a collection of open-source tools built for this purpose, and demonstrate its reliability and accuracy in out-of-sample and LLM-as-a-Judge testing. We extract more than 10 billion data points from more than 155 million online job ads provided by the National Labor Exchange (NLx) Research Hub, including O*NET tasks, occupation codes, tools, and technologies, as well as wages, skills, industry, and more features. We describe the construction of a dataset of occupation, state, and industry level features aggregated by monthly active jobs from 2015 - 2025. We illustrate the potential for research and future uses in education and workforce development.