Evidence that RL teaches genuinely new abilities: compositional skills emerge and transfer across tasks Does RL merely reweight what an LLM already knows—or can it teach genuinely new skills? This paper offers concrete evidence for the latter. Using a controlled, synthetic framework, the authors define “skills” as string transformation
A structured 80k instruction–image corpus spanning 11 domains and 51 subtasks to train unified visual editors Unified models for image generation and editing hit a data ceiling: existing corpora emphasize basic manipulations but miss real‑world complexity. OpenGPT‑4o‑Image tackles this with a hierarchical task taxonomy and automated
ROVER replaces PPO loops with uniform‑policy Q‑values—boosting quality and diversity in math reasoning Popular RLVR methods for LLM reasoning lean on generalized policy iteration (e.g., PPO/GRPO), but suffer instability and diversity collapse. This paper reframes math RLVR as a specialized finite‑horizon MDP with deterministic
A dynamic, self‑paced curriculum that restructures problems to match model ability in RLVR Online RL with Verifiable Rewards (RLVR) has boosted LLM reasoning—but most methods treat all problems equally, wasting effort on solved items and flailing on those beyond current capability. CLPO fixes that with a dynamic pedagogy:
AutoPlay explores environments to synthesize diverse, verifiable tasks—fueling stronger UI agents Agent post‑training hinges on rich, grounded tasks—but human‑curated datasets are costly and often shallow. AutoPlay introduces a scalable alternative: let an MLLM‑based explorer systematically probe interactive environments, discover states and affordances, and then synthesize
A dual‑track “brain‑mouth” design for unified multimodal understanding and expressive, streaming voice MGM‑Omni unifies multimodal reasoning with long‑horizon, controllable speech generation in a single Omni LLM. Rather than bolt speech onto a text reasoning core, it uses a dual‑track, token‑based “brain‑mouth” architecture that
From tabula rasa to transparent mastery: an LLM agent that induces rules and writes its own playbook Most deep RL agents master games by amassing opaque experience. CEL—Cogito, ergo ludo—takes a different path: it learns to play by reasoning and planning. Powered by a Large Language Model, CEL
DataMind: a scalable data synthesis and agent training recipe powering open generalist analytics agents Generalist data‑analytic agents promise automated, end‑to‑end insights across messy, large, and diverse files. But open models often falter on long‑horizon, code‑based reasoning and multi‑turn tool use. DataMind introduces a full
RL from Human Interaction links long‑term personas to turn‑level preferences for scalable, personalized alignment Most alignment today leans on expert‑crafted datasets and static preferences. This paper proposes Reinforcement Learning from Human Interaction (RLHI): learning directly from in‑the‑wild user conversations. Instead of treating chats as fixed
High-quality data over sheer quantity: sub‑billion models that reason competitively with only ~2T core tokens Do we really need tens of trillions of tokens to unlock reasoning in small models? MobileLLM‑R1 argues no. By carefully selecting and resampling about ~2T tokens of high-quality open data, then pretraining to
Evidence that RL teaches genuinely new, compositional skills—beyond mere reweighting. Does RL truly endow LLMs with new capabilities—or just reweight what’s already there? This study provides concrete evidence for genuine skill acquisition via composition. Using a controlled synthetic setup, a “skill” is defined as computing a string
80k instruction–image pairs across 11 domains and 51 subtasks, built with systematic taxonomy. Unified multimodal models stumble when their training data underrepresents real-world complexity. OpenGPT-4o-Image answers with a large-scale dataset constructed via a hierarchical task taxonomy and automated generation pipeline. Beyond basics like text rendering and style transfer, it
ROVER: a minimalist RLVR method that recovers optimal actions from a random policy’s Q-values. LLM reasoning with RL and verifiable rewards often leans on heavy policy-optimization loops (e.g., PPO/GRPO) plagued by instability and diversity collapse. This paper reframes math reasoning as a specialized finite-horizon MDP with deterministic
A dynamic online curriculum that reshapes problems to match—and grow—model ability. Online RL with verifiable rewards (RLVR) is a leading recipe for boosting LLM reasoning, but most methods treat all problems equally—wasting time on what’s already mastered and offering little guidance on what’s just out
AutoPlay explores apps first—then synthesizes diverse, verifiable tasks at scale. Training capable UI agents is bottlenecked by the scarcity of high-quality, grounded downstream tasks. AutoPlay tackles this by explicitly exploring interactive environments to discover capabilities and states before generating tasks. The result: diverse, executable, and verifiable tasks that reflect
A dual-track “brain–mouth” LLM for omnimodal understanding and low-latency, expressive speech. MGM-Omni introduces a unified Omni LLM that cleanly decouples multimodal reasoning from real-time speech generation. Its dual-track, token-based “brain–mouth” architecture enables efficient cross-modal interaction while delivering streaming, low-latency speech that preserves voice identity over long horizons. On
CEL learns rules and strategies from scratch by reflecting on its own trajectories. Instead of memorizing behaviors from vast experience, CEL (Cogito, ergo ludo) learns to play by reasoning and planning in language. Starting tabula rasa (knowing only the action set), the agent interacts with an environment, then reflects on
DataMind: a scalable recipe and dataset to train open, code-capable data-analytic agents. Open-source data-analytic agents lag behind proprietary systems due to data scarcity, brittle training, and unstable multi-turn code execution. DataMind addresses these gaps with a full-stack recipe—data synthesis, curriculum, filtering, objectives, and rollout—to train generalist agents that
RLHI taps in-the-wild conversations to align models to real user preferences and personas. Most conversational models are aligned with expert-annotated feedback—effective yet limited in scale and personalization. This work proposes Reinforcement Learning from Human Interaction (RLHI), a paradigm that learns directly from organic user conversations, connecting long-term personas to
Sub-billion LLMs that reason well—trained on far less data, with a fully open recipe. Do small models need massive corpora to reason? MobileLLM-R1 makes a compelling case that they don’t. Challenging the assumption that advanced chain-of-thought capabilities require >10T tokens, the authors carefully curate and resample open-source
RL with verifiable rewards that define caption quality by downstream utility Supervised captioning has hit a ceiling: it’s costly and tends to memorize specific answers. CapRL reframes the objective with Reinforcement Learning and verifiable rewards (RLVR) for open-ended captioning. The key insight: a high-quality caption should be useful—it
A 10k-example benchmark testing audio understanding, speech, and visual grounding Voice-first assistants are advancing fast, but current tests don’t fully capture what they must do: listen, speak, and see. VoiceAssistant-Eval fills that gap with 10,497 curated examples across 13 task categories—spanning natural sounds, music, spoken dialogue (listening)
Edge-aware attention plus line-search correction for fast, accurate AC power flow Power systems planners need solvers that are both fast and faithful to physics. This work advances Physics-Informed GNNs with two key innovations. First, an edge-aware attention mechanism injects line physics via per-edge biases, capturing grid anisotropy that standard MLP-based
An end-to-end, lossless-feel FP8 recipe that speeds LLM training for reasoning Can we train reasoning-strong LLMs faster and cheaper—without sacrificing accuracy? InfiR2 answers with a practical, open FP8 recipe spanning continual pretraining and supervised fine-tuning. The approach uses a fine-grained, hybrid-granularity quantization strategy to preserve numerical fidelity where it