Reinforcement learning research, RLHF/RLVR developments, agent training — from arXiv cs.LG/cs.AI, HuggingFace, Latent Space, Interconnects, and HN.
Auto-curated: highest-signal RL items from across sources.
Microsoft open-sources Flint, a domain-specific language for declaratively describing visualizations that an agent should generate. Compiles to Vega-Lite / Observable / PowerBI targets.
GitLost researchers show that prompt-injection in a public repo can steer GitHub's AI agent into exfiltrating secrets from private repos the agent has access to. Disclosure timeline and a hardening checklist included.
Databricks engineers benchmark leading coding agents against a real, multi-million-LOC monorepo with a reproducible task suite. Numbers on the agents' actual hit rate on production code, not synthetic tasks.
Adds a dual-latent short/long-term memory module to a VLA policy, lifting success on long-horizon manipulation by ~12 pts over a single-memory baseline. Code and checkpoints released.
RoboDojo unifies sim and real evaluations of generalist robot policies across 12 manipulation tasks with matched protocols, letting researchers report sim-to-real deltas directly. Aims to fix the long-standing gap between leaderboard numbers and lab results.
Reinforcement learning papers from arXiv cs.LG, cs.AI, and Hugging Face Daily Papers.
A deep learning framework for structure-property prediction that ships with dataset cards, error bars, and per-domain baselines. Targets materials science but the transparency tooling generalizes.
Adds a dual-latent short/long-term memory module to a VLA policy, lifting success on long-horizon manipulation by ~12 pts over a single-memory baseline. Code and checkpoints released.
Scales a MoE video pretraining stack to 7B active params and shows the embodied-intelligence downstream suite (CALVIN, SimplerEnv) keeps improving past dense baselines. Useful data point for embodied pretraining economics.
An interactive world model that supports long-horizon, open-vocabulary object interaction. Marketed as a sandbox for training and evaluating agentic policies without a real environment.
RoboDojo unifies sim and real evaluations of generalist robot policies across 12 manipulation tasks with matched protocols, letting researchers report sim-to-real deltas directly. Aims to fix the long-standing gap between leaderboard numbers and lab results.
Uses an LLM-driven outer loop to search over embodied-agent architectures (perception, memory, control) and reports designs that beat hand-tuned baselines on a small manipulation suite. Early but a real step toward agent-design automation.
A 27 km² photorealistic testbed of a real city, scanned and reconstructed for embodied agents and spatial-reasoning benchmarks. Useful for sim-to-real and large-scale world-model work.
Trajectory-review tool that grades coding-agent runs against production telemetry rather than synthetic tests, surfacing failure modes that pass benchmarks but break in real repos. Open-sourced with a public leaderboard.
OmniTacTune is a policy-agnostic RL layer that uses tactile-feedback residuals to adapt visual policies in the real world, no fine-tuning required. Reports a 30% jump in success on out-of-distribution manipulation tasks.
Diagnoses why world-model rollouts collapse on long horizons: the imagined trajectories are kinematically plausible but fail conservation laws and contact dynamics. Proposes a simple energy-consistency loss that recovers most of the gap.
TESSERA v2 scales pixel-wise Earth foundation models to global coverage, with checkpoints and a per-pixel classification head. Useful for remote-sensing and earth-observation downstream tasks.
Token-based dual-view fusion adapts large vision models for breast-cancer classification from mammograms, with weights and a reproducible pipeline. Medical-imaging utility rather than a new paradigm.
Newsletters and HN picks on RL, RLHF, and agent training.
PeTeR post-trains probabilistic circuits to be robust to small input perturbations without losing tractable inference. Across six datasets it recovers most of the accuracy of retrained circuits at a fraction of the compute.
A zero-gradient Riemannian Langevin sampler that uses only function evaluations to draw from multimodal distributions on manifolds. Beats standard MCMC on a suite of synthetic and robotics targets.
Surveys the spectrum of self-improving AI from bounded self-refinement loops to fully autonomous research agents, framing recursive improvement as a control problem with concrete stability conditions. Useful map for anyone working on agentic RL or self-play.
Modal CTO Akshat Bubna on what 'Agent Experience' (AX) means as a counterpart to developer experience (DX), and the infra primitives (sandboxing, persistent state, scheduling) that have to change to support it.
Day 3 dispatch from AIEWF focused on the headlining 'loops' debate, summarizing both sides. Mostly a curation of the day's quotes, useful if you didn't attend.
Vercel Chief of Software Andrew Qu makes the case that agents are a fundamentally new software primitive, not a UX layer over APIs. Short interview but useful framing for anyone building agent infra.
Adobe Principal Scientist Carlos Sanchez on agent-driven web assembly: sites that re-shape themselves per visitor based on inferred intent. Mostly a research-direction talk, light on data.
Paul Bakaus argues that 'skill engineering' — composing narrow agent skills rather than one-shot prompting — is the durable design pattern for production AI products. Concrete examples from Impeccable's product.
Filler digest — 'not much happened today'.
AIEWF floor talk on autoresearch and what it implies for human-in-the-loop product design. Mostly a synthesis of the day's loop-themed sessions, light on new data.
Introspection's Roland Gavrilescu breaks down the autoresearch pattern: agents that grade their own work, update memory, and re-run with a feedback loop. Concrete architecture diagram and failure modes.
Cursor VP of Forward Deployed Engineering on the playbooks enterprise AI deployments actually use, beyond the demo. Useful if you sell into large companies.
Genesis Molecular AI founders on the state of diffusion research outside LLMs: molecular generation, protein design, and the new benchmarks that distinguish real progress from LLM-derived noise.
Warp CEO on 'software factories' as the operational unit for AI-assisted development: small pods of engineers running a long-lived agent fleet. Argues this is where the org chart is heading.
Day 1 dispatch from AIEWF: agent loops dominated the day, with software-factory and forward-deployed engineering framing the operational shape. Light on new data but a clean map of the vocabulary.