ParaTempo prunes 30%, same-origin teacher beats 14B, Hydra-0 cuts motion 90%

Sources

ParaTempo: Efficient Parallel Reasoning via Temporal Confidence huggingface.co

ParaTempo improves parallel reasoning efficiency by using temporal confidence to dynamically prune, retire, and reallocate reasoning branches without synchronization.

Every Coin Has Two Sides: On the Dual Nature of Generalization in On-Policy Distillation of Large Language Models huggingface.co

On-policy distillation transfers reasoning behaviors rather than specific answers, with generalization strongly tied to teacher-student origin alignment and multi-teacher combinations causing capability trade-offs.

Hydra-0: Action Flow for Generalist World Modeling and Control huggingface.co

Hydra-0 uses action flow as a shared visual interface for generalist world modeling and robot control across diverse embodiments and tasks.

Partition the Support, Reconstruct the Residual: Training-Free Sparse Attention for Video Generation and World Models huggingface.co

SparsePR accelerates video generation and world-model transformers without retraining by pairing response-coupled partitioning with probe-fitted residual reconstruction. The method cuts attention-reconstruction error at low executed-pair densities, delivering substantial speedups over dense attention while keeping block-sparse execution friendly to existing video transformer stacks.

CLEAR: Continuous Latent Adapter Routing for Utility-Preserving LLM Safety Alignment huggingface.co

CLEAR tackles the safety-utility tradeoff in aligned LLMs by continuously modulating a low-rank safety adapter through a hidden-state gate. The conditional routing raises HarmBench safety scores on harmful prompts while leaving benign-input utility largely intact, unlike static LoRA merges that degrade helpfulness.

Let’s Scale Step by Step: Compute-Efficient Hyperparameter Transfer for Large-Scale Mixture-of-Experts huggingface.co

A scaling-law framework predicts optimal learning rates for large Mixture-of-Experts models by extrapolating first across width using Maximal Update Parameterization, then across token budgets. The recipe covers Multi-head Latent Attention and the Muon optimizer, cutting pretraining compute otherwise spent on hyperparameter sweeps.

Llama-Mobile: Efficient 2.7-Bit Quantization of VLMs huggingface.co

Llama-Mobile quantizes vision-language models for on-device use with a 2.7-bit-per-parameter format trained on self-generated data. Applied to Llama 3.2 11B Vision Instruct, it shrinks the model to 3.7 GB for Arm CPUs while preserving visual question answering accuracy.

Daedalus-150M: A Convolution-Attention Hybrid Designed for CPU Inference huggingface.co

Daedalus-150M combines sparse attention blocks with short convolutions and 4-bit weights to target CPU inference. The 150M-parameter hybrid outperforms larger conventional transformers on bits-per-byte and downstream benchmarks despite training on far less data, while sustaining longer context under tight memory bandwidth.

Peer-Voted LLM-Agent Stress Tests Find Feed-Induced Lexical Convergence but No Reliable Matched-Exposure Advantage for Distributed Sources huggingface.co

Stress tests of peer-voted feeds populated by synthetic LLM agents show rising TF-IDF lexical similarity across posts, but no reliable opinion capture or matched-exposure advantage for distributed sources. Block-bootstrap analysis held across model families, topics, and adversarial impression injections.

UniSpace: Unified Visual Representation and Scalable Multimodal Modeling huggingface.co

UniSpace reparameterizes a pretrained semantic ViT so a single visual space handles understanding, high-fidelity reconstruction and image generation, removing the usual separate VAE tokenizer. A reconstruction-aware patch embedding plus Mixture-of-Transformer-Experts lets one interface scale across multimodal tasks.

AgentMercury: Your Agent Can Synthesize Verifiable Environments for Business Scenarios at scale huggingface.co

AgentMercury synthesizes scalable executable business environments that serve as generalizable reinforcement learning substrates, improving agent performance across enterprise and out-of-domain reasoning tasks while making environment construction itself learnable.

PhysCaP: Grounding Code-as-Policy Agent with Physics-Informed Exploration huggingface.co

PhysCaP is a physics-informed code-generation agent that actively explores objects to infer hidden physical properties for efficient robotic manipulation.

InfinityEdit: Infinite Video Editing with a Lightweight Edit-Ignition Adapter huggingface.co

InfinityEdit enables continuous, unbounded video editing by using a lightweight adapter with history, temporal causal, and edit cross-attention modules to extend edits into future streaming frames while maintaining stability.

Beyond Correctness: Benchmarking and Aligning Response Behaviors in Hybrid-Thinking MLLMs huggingface.co

Hybrid-thinking multimodal language models suffer from response-pattern misalignment between thinking and non-thinking modes, which is addressed by a diagnostic benchmark and pattern-specific reinforcement learning penalties.

Towards Faithful Simulation of Human Shopping Behavior huggingface.co

RecVerse is a GUI-grounded agent that uses hierarchical memory and trajectory-level reinforcement learning to simulate realistic multi-turn e-commerce shopping sessions.

OmniAssistBench: Assistant-style Interaction Benchmark for Omni-LLMs huggingface.co

OmniAssistBench evaluates real-time interactive video assistants by reverse-engineering multi-turn interaction videos, revealing that current omni-modal models struggle with visual prompts, context retention, and timely responses.

EviRank: Structured Relevance Evidence for Multimodal Image Re-ranking huggingface.co

EviRank reformulates multimodal image re-ranking as semantic constraint satisfaction by parsing queries into structured evidence packages and verifying candidates via rubric scoring and listwise comparison without training.

Graph Engineering in the Era of LLM Agents: From Individual Intelligence to System Intelligence huggingface.co

Graph Engineering organizes multi-agent LLM systems through dynamic graph structures to coordinate specialized agents and manage complex, evolving tasks.

WorldMind: Decoupled Game World Model for State-Aware NPC Behavior huggingface.co

WorldMind decouples game world modeling into state understanding, decision-making, control, and generation layers to enable explicit, state-aware NPC behavior.

Human-Centric Intelligence in the Era of Foundation Models: A Survey huggingface.co

This survey proposes a unified taxonomy and methodological framework for human-centric intelligence across visual, dynamic, and embodied levels within the foundation-model era.

FlavourBench: Ranking Frontier Language Models with Executable Culinary Ground Truth huggingface.co

FlavourBench evaluates language models on culinary portfolio tasks using executable ground truth, statistical rigor, and fully reproducible verification.

Hadith computational science in the age of large language models: a critical narrative review huggingface.co

Hadith computational science is evaluated as an evidence infrastructure challenge requiring integration of transformer and retrieval-based methods with expert validation and provenance.

References

ATM: Any-point Trajectory Modeling (arXiv 2605.23856) arxiv.org

ATM pre-trains a Track Transformer on unlabeled video to predict future paths of arbitrary points, outperforming video pre-training baselines by 80% on average across 130+ language-conditioned tasks; however, side-by-side comparisons show ATM degrades when training and testing embodiments differ significantly, as it struggles to decouple object motion from agent movement.

3DFlowAction GitHub (Hoyyyaard) github.com

3DFlowAction uses a 3D flow world model that predicts the movement of objects rather than robot-specific joint angles… a universal language applicable to humans and diverse robot types alike.

Ego2Robot summary (opentrain.ai) opentrain.ai

Ego2Robot (August 2026) transforms egocentric human videos into robot training data across 15 different morphologies, producing over 18,000 hours of synthetic demonstrations.

NVIDIA Cosmos-Predict/Transfer 2.5 blog (HuggingFace) huggingface.co

NVIDIA provides specialized robotic variants such as Cosmos-Predict2.5-2B/robot/multiview post-trained on three-camera robotic views; fine-tuning employs LoRA and DoRA adapters for viewpoint- or task-specific adaptation.

Alignify world-model comparison alignify.co

1X’s proprietary 1XWM uses a 14B-parameter video diffusion backbone to ‘imagine’ a task before executing it, though inference can take roughly 11 seconds per action… V-JEPA 2 predicts outcomes in an abstract latent space rather than generating pixels, favored for efficiency because it ignores irrelevant visual details.

Twitter/X — lead authors Hongyu Li & Soha Pouya (via twstalker) ww.twstalker.com

Lead researchers emphasized that breaking ‘embodiment silos’ is the critical unlock for scaling Physical AI… community observers noted a gap between the ‘open-source’ announcement and actual data availability; as of late August 2026, the full multi-embodiment dataset was still marked as ‘coming soon’.

EmergentMind — ParaThinker framework overview emergentmind.com

ParaThinker reports accuracy gains of up to 12.3% on 1.5B models and 7.5% on 7B models across mathematical benchmarks like AIME and MATH-500 … with a modest 7.1% latency overhead by leveraging KV-cache reuse.

ACL SRW 2026 paper (SCOUT benchmarking of adaptive SC) aclanthology.org

SAC … terminates paths upon local answer convergence before the final vote. While it reduces token counts, researchers have observed ‘catastrophic drops’ in accuracy (e.g., from 68.6% to 63.2% on Qwen3-8B) because local convergence doesn’t always align with global correctness.

aimodeling.com technical review aimodeling.com

ParaTempo’s accuracy is described as ‘competitive’ rather than ‘lossless’ … it is an inference-time scheduler; it improves the efficiency of finding a solution but does not raise the underlying model’s capability ceiling.

NVIDIA Developer Blog — speculative decoding / tree attention developer.nvidia.com

Tree-based branch pruning in vLLM can yield a 1.4x to 1.6x speedup on models like Llama-3 … however, if the draft model’s acceptance rate falls below approximately 50-60%, the overhead of managing the speculative branches can actually lead to a net performance loss.

OpenReview — SeerSC (advance budget estimation) openreview.net

SeerSC … uses a rapid ‘System 1’ pass to calculate answer entropy before committing to a full reasoning chain … reduces token consumption by up to 47% and latency by 43% by only applying heavy self-consistency checks to complex queries.

vLLM GitHub issue #18327 (dynamic tree topology) github.com

Developers have pushed for more dynamic pruning capabilities, specifically requesting support for dynamic tree topology in methods like EAGLE-3 to adapt to varying prompt complexities.

Thinking Machines Lab blog (Kevin Lu) thinkingmachines.ai

A 0.5B parameter student matches a 1.5B teacher on math reasoning tasks while using 9–30x less compute than traditional RL, with the loss expressible in roughly ten lines of PyTorch as reverse-KL between student and teacher logits on student rollouts.

bdtechtalks — LLM self-distillation tradeoffs bdtechtalks.com

Teachers using ‘privileged information’ to supervise students discourage the expression of uncertainty; suppression of epistemic markers like ‘Wait’ or ‘Let me recheck’ can cause performance drops of up to 40% on out-of-distribution problems.

PACED (ACL Findings 2026) aclanthology.org

Signal-to-noise ratio follows a bell curve p(1-p) relative to student pass-rate and collapses at both 0% and 100% extremes — both easiest and hardest data should be filtered simultaneously, contradicting classical easy-to-hard curriculum learning.

arXiv 2607.15161 — ‘Illusory’ distillation analysis arxiv.org

OPD-trained models show superior avg@K but their pass@K for large K often lags their pre-distillation base versions; previously solvable difficult problems become unsolvable, suggesting capability collapse in favor of teacher-style shortcuts.

alphaXiv commentary on 2608.19408 (concurrent work) alphaxiv.org

Scaling model-generated distillation data can make latent teacher traits more recoverable — subliminal inheritance of behaviors that survive explicit data filtering.

alphaXiv discussion of ‘Every Coin Has Two Sides’ (2608.16647) alphaxiv.org

Origin compatibility gives developers a practical ‘lineage-first’ heuristic; the seesaw effect shows teacher influence in same-origin MOPD cannot be confined by prompt routing because the student’s policy is altered as a whole.