Agent Lightning and LEGO-RL credit the harness; Abra hits 10× Chinchilla

Sources

Agent Lightning v1.0: Towards Harnessed Agentic RL huggingface.co

Agent Lightning v1.0 enables reproducible reinforcement learning for arbitrary agent harnesses, substantially improving coding-agent performance with minimal data and compute.

LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents huggingface.co

LEGO-RL connects native coding-agent harnesses to scalable policy-gradient training via in-process LLM proxying, sandbox orchestration, and integrated monitoring, improving sparse MoE model performance across multiple harnesses.

Abra: Scaling Diffusion Image Training huggingface.co

Scaling laws for text-to-image diffusion models reveal predictable compute-optimal training requiring far more data per parameter than language models, with robust overtraining behavior and universal curve shapes.

Demystifying Agent Skills: Why They Work-Until They Don’t huggingface.co

Skills enhance LLM agents primarily by stabilizing execution through procedural anchoring rather than injecting missing knowledge, though retrieval bottlenecks and brittle assumptions limit their effectiveness.

Harness the Memory: A Holistic Evaluation of Memory Substrates in Memory Agents huggingface.co

Long-horizon agent memory has no universal best design, according to an empirical sweep across dense, sparse, structural, hierarchical, refinement-based and parametric stores. Trade-offs shift by regime, so the authors argue for adaptive substrate routing rather than committing agents to one memory backend.

MathForm: Scaling Mathematical Autoformalization with Knowledge Retrieval and Verification-Guided Refinement huggingface.co

Autoformalization of math into Lean improves when a retrieval planner pulls Mathlib lemmas and compiler diagnostics guide iterative refinement. The pipeline produced FormalVerse, a verified training set, and an 8B model trained with SFT plus RL that leads on Pass@8.

PTXBench: Benchmark and Adapt LLMs for GPU Kernel Optimization with Architecture-specific PTX huggingface.co

GPU kernel optimization at the PTX level exposes uneven LLM performance across GEMM and attention workloads on H100 and B200 hardware. Supervised fine-tuning and repair-conditioned training with a reasoning teacher close only part of the gap between generated kernels and expert baselines.

FreeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution huggingface.co

Mixture-of-Experts serving comes to laptops and workstations through FreeToken, which dynamically places experts and KV state across CPU and GPU based on bandwidth. Runtime memory management and agentic state reuse let open-weight MoE models run locally without dedicated datacenter accelerators.

SkillForge: Self-Distilling Agents for Project-Specific Issue Resolution huggingface.co

Automated issue resolution improves when agents first practice on synthetic bugs generated inside the target repository. SkillForge distills entity-grounded skills tied to specific files and functions, giving the agent a project-specific playbook before it sees any real GitHub issue.

HarnessRisk: A Lifecycle-Oriented Benchmark for Agent Harness Safety huggingface.co

Agent harness security gets scrutinized across the operational lifecycle and inside DeepSeek’s deployment. HarnessRisk maps six phases where configuration flaws let attacks succeed while utility looks intact; the DeepSeek audit uses taint tracking and dual judges to measure indirect prompt injection through text and file channels.

Security Assessment of DeepSeek Harness with A.I.G: Evaluating Resistance to Indirect Prompt Injection huggingface.co

Researchers evaluate indirect prompt injection risks in DeepSeek Harness using controlled taint and dual judges, finding notable success rates across text and file channels and recommending controls between untrusted content and sensitive actions.

From Sequence to Structure: Relational Uncertainty Propagation for LLM Agents huggingface.co

Long-horizon LLM agents get better failure detection when execution is modeled as a directed dependency graph rather than a flat sequence. RUPA propagates step-level uncertainty along those edges, sharpening trajectory-level confidence estimates and flagging errors earlier than sequence-only baselines.

Dynamic Multi-Byte Prediction With Hierarchical Language Models huggingface.co

Multi-byte prediction accelerates byte-level hierarchical language models by generating parallel bytes via variable-length windows and causal attention masking, improving inference speed with minimal quality loss.

Unifying Graph Neural Networks Through a Common Layer Equation huggingface.co

A unified layer equation decomposes graph neural networks into seven components to compare architectures, derive theoretical bounds, and expose design choices linked to oversmoothing and expressivity.

Agentic ESOpt: Fine-Tuning Long-Horizon LLM Agents with Minimal GPU Requirements huggingface.co

Agentic ESOpt uses evolution strategies for scalable full-parameter fine-tuning of long-horizon LLM agents via trajectory-level reward-weighted updates and parameter-context co-evolution.

MoE-ViE: Mixture of Experts Vision Encoder for Efficient Image and Video Understanding huggingface.co

Mixture-of-Experts vision encoders with fine-grained topologies, auxiliary-loss-free balancing, and specialized kernels scale efficiently while outperforming larger dense models on image and video tasks.

The Problem Is the Problem: Towards Scalable Mathematical Discovery huggingface.co

A literature-to-review pipeline automates problem discovery and triage to focus expert review on promising mathematical conjectures within a chosen research direction.

ASI-Bench: At the Dawn of Artificial Superintelligence huggingface.co

Artificial superintelligence (ASI) requires AI to move beyond mastering existing knowledge toward exploring the unknown, creating new knowledge, and turning new ideas into verifiable results. However, the capabilities of today’s AI systems are still largely built on learning, compressing, and applying existing human knowledge. Accordingly, existing benchmarks primarily test whether AI can produce correct answers based on learned knowledge, or whether it can complete tasks under extensive human g

StartupBench: Benchmarking General-Purpose Agents on Market-Validated End-to-End Workflows huggingface.co

StartupBench evaluates end-to-end AI agents on real-world startup workflows and reveals that even top models complete only about 30% of tasks, highlighting gaps in instruction following and domain expertise.

Cross-Model Memory Transfer via Target-Side Reader Adaptation huggingface.co

Cross-model reuse of frozen external memory tables depends primarily on aligning a lightweight target-side reader rather than the memory content alone, enabling reusable knowledge artifacts with optional adaptation.

V-RAE: Rethinking Video Latent Spaces for Generation huggingface.co

V-RAE constructs semantically organized video latents from frozen vision representations to improve generation quality, convergence speed, and predictive modeling.

Energy-Guided Flow Matching huggingface.co

Energy-Guided Flow Matching improves generative quality by progressively revealing high-frequency details through a moving endpoint and adaptive scheduling, reducing training cost and achieving state-of-the-art FID scores.

PixRestore: Unified Image Restoration via Pixel Diffusion Transformer huggingface.co

PixRestore is a compact, VAE-free pixel-space diffusion transformer trained from scratch for unified image restoration, using flow matching on patchified pixels, DINO-based reliability-guided feature fusion, and adversarial fine-tuning to a one-step generator for efficient high-fidelity inference.

Personalized Auto-Research: Towards a True AI Co-Scientist huggingface.co

The paper introduces personalized auto-research, a framework that conditions AI-driven hypothesis generation, experimentation, and writing on individual researcher representations to avoid generic outputs.

From Corpora to Co-Evolving Capabilities: Capability-Centric Data Design for Generalist Image Generation huggingface.co

A capability-driven data infrastructure with curriculum scheduling and specialized data engines trains large multimodal diffusion models on curated heterogeneous supervision for diverse generative tasks.

DiSCO: Defending text-to-image generation through distribution-guided contrastive prompt optimization huggingface.co

DiSCO is a black-box, zero-shot prompt-level defense that uses distribution-guided suffix expansion and contrastive scoring to reduce harmful image generation without altering the model.

aDSL: Agentic 3D Creation via Joint Agent-Program Design huggingface.co

A co-designed domain-specific language and multi-agent system improve LLM-driven 3D program synthesis by using relational operators and iterative execution feedback.

AVA-Encoder: Towards Agent-Native Video Representation Learning huggingface.co

AVA-Encoder learns structured video representations via agentic auto-encoding using knowledge graphs and textual-gradient optimization to improve cinematic video generation and reasoning.

CoinVE-200K: A Large-Scale High-Quality Dataset for Compositional Instruction-Guided Video Editing huggingface.co

A new dataset, benchmark, and 22B model enable compositional instruction-guided video editing with multi-region attention and temporal coherence.

EDITBRIDGE: Towards Faithful and Efficient Ultra-High-Resolution Image Editing huggingface.co

EditBridge enables efficient ultra high-resolution image editing via a diffusion bridge that translates low-resolution edits to high-resolution outputs while preserving source details through sparse attention.

Embodied-Navigator: Point, Think, Memorize, and Align for Efficient Navigation huggingface.co

TAMP-Nav improves embodied navigation by aligning vision-language models with 2D visual prompting, selective reasoning with compressed memory, and dense policy optimization.

CardioState-JEPA: Delay-Aware Cross-Modal Learning of a Shared Cardiac Representation huggingface.co

CardioState-JEPA learns a unified cardiac representation across ECG, PPG, and PCG by predicting masked latent physiological states with cross-modal delay alignment, improving downstream classification across all three modalities.

GS-Voxel: Fitting-Free Structured Latents for Large-Scale 3DGS Generation huggingface.co

GS-Voxel converts unstructured 3D Gaussian reconstructions into sparse structured latents to enable scalable aerial scene generation via flow models.

References

The Moonlight review of LEGO-RL themoonlight.io

LEGO-RL demonstrated consistent improvements across all supported harnesses: OpenHands SDK rose from 64.0% to 70.4%, and Claude Code improved from 62.4% to 68.2%… Rollout Routing Replay (R3)… records the specific expert indices chosen during the initial generation and forces the trainer to use those same experts during the backward pass.

vLLM blog, ‘No More Retokenization Drift’ (Oct 22, 2025) vllm.ai

The integration addressed this by modifying the vLLM API to return exact token IDs alongside text, ensuring that the training data perfectly matches the inference-time execution.

Hacker News discussion news.ycombinator.com

Several developers compared the framework to a ‘worse DSPy,’ arguing that while it ‘plumbs things together,’ it does not necessarily solve fundamental algorithmic issues like sparse rewards or partial observability… some users describing the README as ‘bizarre’ and questioning the presentation quality of a multi-trillion-dollar company.

openlm.ai SWE-bench leaderboard notes openlm.ai

Reports suggest that some agents achieve high scores by utilizing git log commands to ‘peek’ at future repository states, effectively finding the solution in the version history rather than reasoning through the bug… many independent leaderboards now prioritize ‘bash-only’ evaluations using the mini-SWE-agent harness.

Microsoft Research blog (original Agent Lightning announcement) microsoft.com

The original August 2025 paper introduced Training-Agent Disaggregation (TA Disaggregation), which decoupled the agent’s execution from the reinforcement learning training engine… v1.0 release (August 2026) centers on ‘Harnessed Agentic RL’ [where] the trainer acts as a silent observer of API request-response pairs.

GitHub microsoft/agent-lightning issues + adoption github.com

Agent Lightning has approximately 17,700 stars and 1,600 forks… GitHub issues reveal bugs in multimodal RL, where vision features are sometimes ‘silently dropped’ during training… setup remains complex, often requiring Kubernetes or specific GPU clusters (e.g., 8× A100/H100).

AI Weekly editors’ blog aiweekly.co

found first: 65.7% of agent skill gains trace to procedural anchoring in 8,135 [trials]… explicit knowledge injection accounts for only 4.5% of successful skill usage

TechPulse Facebook post / HN-adjacent discussion facebook.com

roughly 76 [of nearly 4,000 public skills] contained malicious code… skills are ‘closer to npm packages than README files’

AI Accelerator Institute — ‘Your AI agents’ skills are lying to you’ aiacceleratorinstitute.com

SkillsBench… human-curated skills improved task success by an average of 16.2 percentage points, [while] skills generated by LLMs themselves provided negligible or even negative benefits

The Decoder — Voyager coverage the-decoder.com

Voyager stores ‘how-to’ knowledge as JavaScript functions… successful programs are indexed via embedding vectors of their descriptions for future retrieval and composition

Thoughtworks — ‘Harnessing agent semantic reliability at scale’ thoughtworks.com

Frameworks like Claude Code now implement PreToolUse hooks that sit outside the LLM’s context, allowing the system to deterministically block an invocation if a hard precondition… is not met

Dimitrij Pankratz on Medium — ‘A trust problem’ medium.com

Because the agent—not a hardcoded path—decides when to call a script based on natural language instructions, security audits are difficult. An exploit could remain dormant during standard scans and only trigger under specific conversational states

AI Weekly summary aiweekly.co

compute optimality for diffusion transformers occurs at approximately 200 image tokens per parameter… ten times the 20-token-per-parameter ratio established by the Chinchilla scaling laws

OpenTrain.ai review opentrain.ai

there is currently no verified maintained repository or third-party reproduction of the benchmarks

SambaNova blog on inference economics sambanova.ai

inference costs can account for 80-90% of a model’s total lifecycle expense… it is possible to reduce inference compute by approximately one order of magnitude while maintaining performance if the training compute is increased by 1-2 OOMs

alphaXiv discussion of Amazon ‘Scalability of Diffusion-based T2I’ paper alphaxiv.org

increasing the number of transformer blocks is more parameter-efficient than simply increasing channel counts… a U-Net variant that is 45% smaller and 28% faster than the widely used SDXL

arXiv 2505.15270 – Scaling Diffusion Transformers Efficiently via µP arxiv.org

a DiT-XL-2 model using µP reached convergence 2.9 times faster than its standard-parameterized baseline… MMDiT architectures (up to 18B parameters)… only 3% of the tuning budget typically consumed by human experts

The Neuron AI daily digest theneuron.ai

a frozen VAE acts as a ‘latent bottleneck’ that limits fine-grained detail and caps the potential of the larger model, regardless of how many tokens it processes