Sources

Feyospace-v1: How the Cyber Mercury Seven Trained Frontier Cyber Models huggingface.co

A data-centric framework with specialized systems for reasoning analysis, cost reduction, and execution verification enables small teams to train open-weight cyber agents that achieve top-tier performance on benchmark suites.

Occamy-1.0: Open Pareto-frontier 35B Intelligence for Co-work huggingface.co

Occamy-1.0 is a cost-efficient co-work agent model that achieves strong performance on complex multi-step tasks while maintaining broad agentic capabilities at lower cost.

ReactHuman: A Physics-Grounded Benchmark for Human-Like Reactive Decision-Making in Embodied Multimodal LLMs huggingface.co

ReactHuman is a physics-grounded benchmark that evaluates multimodal large language models as decision cores for simulated humanoid robots facing sudden household hazards, revealing persistent reactive safety failures.

DataFlex-RL: An Evaluation Platform for RLVR Data Policies huggingface.co

Uniform sampling matches or beats adaptive rollout selection, reweighting, and domain mixing on RLVR training, according to DataFlex-RL benchmarks across math, logic, and science. Tests on Qwen2.5-7B-Base and Llama-3.1-8B-Base include GPQA-Diamond, questioning the payoff of complex data curation for GRPO-style training.

Ambient @ EgoLongQA 2026: Distilling Long-Video perception into a Sub-2B Model huggingface.co

First-place finishes in both EgoLongQA and EgoProactive tracks came from a 2B vision-language model distilled from a tool-using agent, with its multilingual embedding table pruned to hit parameter limits. The proactive-assistance entry recasts intervention timing as single-token classification supervised by a video agent.

Ambient @ EgoProactive 2026 : Proactive Egocentric Assistance with Visually Grounded Supervision huggingface.co

The approach reformulates intervention timing as single-token classification and uses a video agent for visual supervision to improve wearable assistant decision-making.

StepAudio 3 Gen Technical Report huggingface.co

A single discrete autoregressive model handles text-to-speech, voice design, sound effects, and music by sharing a residual vector quantization code space across modalities. StepAudio 3 Gen pairs a causal Transformer with an RVQ Adaptor and interference-aware progressive pretraining to avoid task collisions.

SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking huggingface.co

Sparse attention gains a learned context ranker by inserting continuous softmax gates during training and swapping in Top-K selection at inference. Tencent Hunyuan’s SAS ships a Triton kernel compatible with FlashAttention, improving long-context accuracy under tight token budgets versus dense-attention distillation baselines.

Pelican-Sim 1.0: A General World Model Simulator for Embodied Intelligence huggingface.co

Predicting future frames from visual context and robot actions, Pelican-Sim 1.0 uses unified URDF-based action representations, action-visual injection, and a sparse mixture-of-experts backbone. Few-step distillation and causal adaptation speed rollouts, targeting controllable video generation across varied embodiments and manipulation tasks.

PLC-DPO: Posterior Label Correction in Noisy and Ambiguous Preference Optimization huggingface.co

Noisy or ambiguous pairwise labels get routed into clean, flipped, or tied buckets by PLC-DPO using calibrated policy-reference margins during preference optimization. The posterior correction step targets a common failure mode of vanilla DPO, where mislabeled pairs push the policy in the wrong direction.

Towards a Deterministic Math Solver for Clinical Language Models huggingface.co

Delegating arithmetic to a restricted Python executor improves clinical calculator scores on MedCalc-Bench Verified for larger open-weight models but not reliably for smaller ones. The program-solve interface does not substitute for verified formulas or accurate variable extraction, limiting deterministic gains in clinical LLM pipelines.

Breaking the Vision-Action Shortcut: Latent Interface Training for Generalizable Robotics Foundation Models huggingface.co

LIT improves robot action generalization by first training pose-conditioned action priors without images, then constraining visual inputs through a pose-supervised latent interface that preserves spatial goal information.

SNAP3D: Physically Grounded 3D Parts for Assembly from a Single Image huggingface.co

A physics-guided framework improves part-aware 3D generation by resolving inter-part penetration, recovering contact graphs, and refining parameterized connectors via simulation feedback to ensure stable, physically valid assemblies.

Studying Without a Syllabus: Task-Agnostic Environment Preprocessing huggingface.co

An agent can explore unfamiliar environments without task-specific guidance to build reusable artifacts that reduce later inference costs, though larger study budgets do not always improve results.

TRACE: Trajectory-robust Admission with Evidence Ordering for Efficient GUI Agents huggingface.co

TRACE is a training-free framework that ranks visual evidence by future utility and diversity, reserves native tokens for spatial coverage, and contracts retired frames to reduce latency and memory in GUI agents.

Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation huggingface.co

Benchmark Radar is a searchable living database and discovery engine for AI evaluation benchmarks that aggregates sources, score histories, and evidence to support benchmark selection and comparison.

ActionSplice: In-Flight Action Editing for Interactive World Models huggingface.co

ActionSplice introduces counterfactual state transport to splice revised actions into chunk-autoregressive video world models without replaying completed evaluations, improving fidelity and speed.

How Far Can Synthetic Data Take Thai OCR? huggingface.co

A Thai OCR model trained solely on synthetic documents achieves strong real-world performance by isolating key transfer factors like typography, layout, and handwriting glyphs.

Building and Evaluating Fixed-Voice Thai TTS from Synthetic Speech huggingface.co

A compact Thai text-to-speech model is distilled from a large voice-cloning teacher using synthetic data from brief voice references, achieving strong on-device accuracy and prosody with minimal errors.

COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization huggingface.co

COBRA-Skills improves LLM agent skill optimization by using contextual-bandit prioritization and evidence-based evolution to cut evaluation costs while maintaining high performance.

Beyond Top-k Skill Retrieval: Diversity-Aware Skill Routing for LLM Agents huggingface.co

Diverse Skill Routing improves LLM agent skill selection by balancing relevance with non-redundancy via a determinantal point process, boosting multi-skill coverage.

Competence-Gated Pooling of Language Models and Priors for Event Forecasting huggingface.co

A competence gate selectively integrates language model forecasts by estimating domain-level marginal value over external predictions, improving hybrid forecasting accuracy.

ReMoMask-2: Latent Retrieval-Augmented Masked Motion Generation huggingface.co

ReMoMask improves text-to-motion generation via hierarchical contrastive retrieval and topology-aware fusion, while ReMoMask-2 eliminates representation gaps by embedding retrieval directly into the generator’s latent space.

Online Learning with LLM Experts from Limited Feedback huggingface.co

Adaptive routing of prompts to LLM experts is formulated as a contextual bandit problem with limited feedback, yielding algorithms with sublinear regret bounds and effective routing strategies.

MobileVLA-R1 2.0: RL-Enhanced Reasoning for Mobile Robot Control huggingface.co

MobileVLA-R1 2.0 couples structured reasoning with robot control via chain-of-thought alignment and reinforcement learning to improve long-horizon mobile robot instruction following.

Feature Recovery for Object Understanding After Irreversible Fire Damage huggingface.co

TRACE is a benchmark for post-fire object understanding that introduces a feature recovery module to restore degraded representations and improve detection and vision-language tasks under severe physical damage.

References

Dark Reading — US government on Chinese AI distillation darkreading.com

Anthropic released a 154-page report identifying and blocking large-scale distillation by seven Chinese labs… redefined these activities as ‘illicit distillation attacks,’ distinguishing them from legitimate model compression.

Industrial Cyber — Dream’s Hero on CyberGym industrialcyber.co

Dream’s Hero—built on GLM-5.2 open weights—has achieved scores as high as 96.6% by orchestrating multiple specialized agents, versus Feyospace-s1’s 63.24%.

UC Berkeley RDI blog — CyberGym methodology rdi.berkeley.edu

Experts note that while top-tier agents now consistently achieve high scores, modest point differences may not represent meaningful capability gaps due to the stochastic nature of agent runs; any performance gap smaller than 15 points is ‘indistinguishable from measurement error.’

ogwilliam.com — ‘Hydra Clusters’ analysis blog.ogwilliam.com

These services often rely on ‘Hydra Clusters’—sprawling networks of thousands of fraudulent accounts created using stolen credentials or fake credit card numbers to circumvent rate limits and geographic blocks.

Lawfare — Responding to AI distillation without panic lawfaremedia.org

Dissenters note that if AI outputs are not copyrightable, distillation may simply be a form of functional imitation rather than theft… prominent figures have argued that using other AIs to validate one’s own is ‘standard practice.’

AI Weekly editors’ blog — Feyospace safety regression aiweekly.co

Since they prioritized capability and helpfulness without a separate alignment stage, successful execution of harmful cyber requests climbed above 86%, a ‘hard regression’ in safety alignment.

AI Weekly benchmark alert aiweekly.co

Occamy scored 82.20 on Claw-Eval, edging GPT-5.6 Sol’s 81.80… but only 49.16 on WildClawBench versus GPT-5.6 Sol’s 67.20, and trailed significantly on Terminal-Bench 2.1 (59.0 vs 88.80).

VentureBeat (Alibaba Metis coverage) venturebeat.com

HDPO reduced redundant tool invocations from 98% to just 2% while simultaneously improving overall reasoning accuracy.

cctest.ai analysis cctest.ai

Because the model’s efficiency claims rely on the authors’ own frozen pricing protocol and self-selected benchmark mix (Claw-Eval, WildClawBench, AutomationBench, GDPval), the Pareto-frontier result may not generalize to different API structures or hardware.

claw-eval GitHub repo github.com

Switching the agent harness alone could shift a model’s score by up to 18 points, making it difficult to attribute performance gains solely to the underlying LLM.

Unite.AI on Accio unite.ai

In a 107-task e-commerce benchmark, Accio completed tasks at an estimated 50% lower cost than OpenAI’s Codex and Anthropic’s Claude Code.

Hugging Face Accio-Lab/occamy-1.0-GGUF (community quants) huggingface.co

Community-provided GGUF, FP8, NVFP4 and MLX builds enable execution on Apple Silicon and 24GB VRAM GPUs at 80–90 tokens/sec on consumer dual-GPU setups.

Medium (Sai Manideep) — Genesis engine review citing Stone Tao critique medium.com

Corrected benchmarks using standard accuracy settings suggest that Genesis may actually be 3–10x slower than alternative GPU simulators like ManiSkill… headline-grabbing benchmark numbers were achieved by using ‘the fastest physics solver configurations,’ such as setting physics substeps to 1, which significantly degrades simulation accuracy.

cctest.ai review of ReactHuman cctest.ai

Larger, more sophisticated models often perform worse on safety metrics than their smaller counterparts… ‘frozen’ default; when faced with complex physical events, they frequently fail to move at all, accounting for over 60% of safety failures in some tests.

TheMoonlight.io review — methodological critique themoonlight.io

Biological brains do not ‘freeze’ their synapses or wait for a discrete t0; instead, they rely on continuous adaptation and predictive processing… open-loop predictions often diverge from real-world performance because they do not account for how a model’s own policy might induce novel, unobserved states.

Embodied-AI-Daily GitHub (Whole-Body-Control survey) github.com

A leading architectural trend is the decoupling of the ‘brain’ (MLLM planner) from the ‘spine’ (RL-based motor controller) and the ‘body’… a pre-trained RL controller handles the high-dimensional contact dynamics and balance required for the humanoid torso and limbs.

alphaXiv discussion of ReactHuman alphaxiv.org

Discussions… emphasize that while models can articulate physics principles in text, they struggle to apply them in real-time ‘closed-loop’ environments… developers on Hacker News highlighted the ‘adversarial objects’ as a particularly effective ‘vibe check’ for embodied AI.

Sources

References

Jack Sun, writing.