Co-evolving a policy LLM and a generative process reward model for OR Solving Operations Research problems with LLMs demands more than final-answer rewards. StepORLM introduces generative process supervision that evaluates the entire modeling and reasoning pipeline. At its core is a co-evolution loop: a policy model learns to solve OR
C-FREE fuses 2D graphs with 3D conformers using ego-nets—no negatives required High-quality molecular representations often require large labeled datasets or fragile contrastive schemes. C-FREE offers a simpler, more powerful path: contrast-free, multimodal pretraining on both 2D topology and ensembles of 3D conformers. The core idea is to predict subgraph
Auto-generating, routing, auditing, and evolving multi-agent systems as a DAG pyramid Building effective LLM agents typically demands bespoke workflows, prompts, and tuning—a bottleneck to scale. InfiAgent proposes a pyramid-like DAG framework that automates agent creation, coordination, and continuous improvement across “infinite” scenarios. Its generalized “agent-as-a-tool” mechanism decomposes complex tasks
Diagnosing LLM reasoning failures via geometric deviation from a manifold of success Why do large language models succeed on some reasoning chains and fail on others? REMA introduces a compelling geometric perspective: the Reasoning Manifold—an emergent, low-dimensional structure formed by internal representations associated with correct multi-step reasoning. Errors, in
Clinically grounded guardrails that make AI draft responses safer and more complete As patient-portal messaging scales, clinicians need AI assistance that is accurate, empathetic, and workflow-aware. This work introduces a practical blueprint for building such guardrails. First, the authors craft a clinically grounded error ontology—spanning 5 domains and 59
A training-free VLM approach that turns language into waypoint-guided UAV control See, Point, Fly (SPF) reimagines how unmanned aerial vehicles follow natural-language commands—without any additional training. Instead of treating action prediction as text generation, SPF reframes aerial vision-and-language navigation as 2D spatial grounding. The system decomposes open-ended instructions into
Objective rewards for a subjective task: captions judged by how well they enable QA. Supervised captioning often memorizes ground-truth phrasing and struggles to generalize. CapRL brings Reinforcement Learning with Verifiable Rewards (RLVR) to image captioning by redefining “good” captions in terms of utility: can a non-visual LLM answer questions about
10,497 examples, 13 tasks: a holistic yardstick for voice-first multimodal assistants. Voice assistants are rapidly evolving into multimodal agents that must hear, speak, and see. Yet evaluation has lagged behind capability. VoiceAssistant-Eval fills this gap with a comprehensive benchmark of 10,497 curated examples across 13 task categories, spanning
Fast, accurate AC power flow with edge-aware physics and an inference-time correction loop. For grid planners and operators, AC power-flow solves must be both fast and precise. This physics-informed GNN (PIGNN-Attn-LS) advances both fronts by baking line physics into the model via edge-aware attention and reintroducing an operative decrease criterion
An end-to-end FP8 recipe that’s stable, lossless vs BF16, and measurably faster for reasoning LLMs. Training frontier LLMs is costly. FP8 promises efficiency, but real-world adoption has lacked a robust, open recipe. InfiR2 fills that gap with an end-to-end FP8 training methodology spanning continual pretraining and supervised fine-tuning, underpinned
Co-evolving policy and process reward with solver-grounded feedback for OR tasks. LLMs are promising solvers for Operations Research, yet two pitfalls persist: outcome-only rewards misassign credit, and discriminative process supervision often misses dependencies across modeling steps. StepORLM addresses both with a self-evolving loop that couples a policy model with a
C-FREE unifies 2D topology and 3D conformers—no negatives, no positional encodings, no heavy preprocessing. Molecular representation learning often relies on contrastive schemes, hand-crafted augmentations, or complex generative objectives—frequently ignoring the rich 3D geometry that governs chemistry. C-FREE (Contrast-Free Representation learning on Ego-nets) offers a simpler, stronger path: learn