## The big picture

## Architectural breakthroughs

**DiffusionART: Aggregation-First Reward Tilting**
The authors propose DiffusionART, which aggregates rewards for each prompt into a shared direction for reverse-process states, eliminating the destabilizing forward-reverse consistency force inherent in NFT-style objectives. By adding a trust-region-masked second update for drift-adaptive restoration, the method achieves stable and efficient online reinforcement learning for diffusion models without requiring carefully tuned EMA off-policy training.
[HF Daily Papers](https://huggingface.co/papers/2610.12572)

**Cryptographic backdoors for hidden capabilities**
The authors construct backdoor attacks that embed encrypted neural networks within host model weights, decrypting and executing them only upon receiving a secret trigger. These "digital lockers" are provably hard to elicit or interpret under cryptographic assumptions, even with full access to model weights, demonstrating that current evaluation protocols cannot guarantee the absence of hidden capabilities.
[arXiv](https://arxiv.org/abs/2610.13928v1)

**Run-to-run variance in semantic queries**
The authors measure non-determinism in semantic query engines by re-executing workloads from SemBench, LRO-Bench, and UDA-Bench, finding that 0.07% to 10.2% of rows change outputs at temperature 0. They show that these small per-row variances amplify exponentially in complex plans, causing up to two-thirds of final answers to differ between executions, undermining the reliability of current benchmarking practices.
[arXiv](https://arxiv.org/abs/2610.13823v1)

**Hide and Seek: Black-box concept attribution**
The authors propose Hide and Seek (HnS), a method for concept-level attribution in black-box vision models that requires no access to weights or labeled concept datasets. HnS retrieves concept-aligned images using a vision-language model, masks localized regions, and measures attribution from the change in the model's output, enabling interpretation of a wide range of opaque models.
[arXiv](https://arxiv.org/abs/2610.13935v1)

**SynthSelf: Continual learning for small models**
The authors introduce Synthetic Self-Distillation (SynthSelf), which enables continual learning for small, on-device models by outsourcing in-distribution rewriting to a minimally larger, same-family aligner. This approach addresses the failure of standard self-distillation fine-tuning in smaller models by ensuring in-distribution learning, coupled with off-policy supervision and rejection sampling.
[HF Daily Papers](https://huggingface.co/papers/2610.13113)

**DCOCO: Contextual correction for speculative decoding**
The authors introduce DCOCO, a lightweight mechanism that restores intra-block dependencies in parallel diffusion-based drafters for speculative decoding. By correcting tokens without committing to top predictions, DCOCO preserves the parallelism of diffusion drafters while mitigating the loss of information from preceding draft tokens.
[HF Daily Papers](https://huggingface.co/papers/2610.12570)

**Accepting-path decoding for semantic parsing**
The authors propose accepting-path decoding (APD), which maintains a verified completion within a fixed output-length limit during neural semantic parsing. APD guarantees an accepted output without retracting tokens, significantly improving Smatch++ scores over local masking under tight length constraints.
[arXiv](https://arxiv.org/abs/2610.13896v1)

**S-Space: Spatial workspace in MLLMs**
The authors uncover S-Space, a linear representation subspace in multimodal language models where object-token representations encode continuous coordinates for horizontal, vertical, and distance axes. They demonstrate that S-Space functions as an internal spatial workspace, remaining readable even under non-spatial questions and supporting accurate spatial judgments.
[HF Daily Papers](https://huggingface.co/papers/2610.13527)

**Task-dependent attention head roles**
The authors analyze how shared attention heads in large language models redistribute their influence across different tasks, showing that task changes alter which facts the model reads even when instructions remain fixed. This work provides a network view of language-model computation, revealing that fixed components adapt their roles dynamically based on task context.
[arXiv](https://arxiv.org/abs/2610.13895v1)

**ReHarness: Recurrent harness for agentic continual learning**
The authors propose ReHarness, a framework for online continual learning in agents that complements forward information flow with backward revision. After each instance, the system leverages the resulting trajectory as feedback to revise the reflections retrieved for that instance, propagating feedback backward through the textual harness state to better inform future interactions.
[HF Daily Papers](https://huggingface.co/papers/2610.12668)

**PreSTO: Predictive subtree prefetching**
The authors propose PreSTO, which accelerates PowerMH sampling for LLM reasoning by prefetching future proposals as a tree of accept/reject decisions. PreSTO generates proposals for a prefetched subtree in one batch, then follows the realized path with the original Metropolis-Hastings acceptance rule, provably leaving the transition kernel unchanged.
[HF Daily Papers](https://huggingface.co/papers/2610.13171)

**SpatialASM: Active spatial mentalization**
The authors introduce SpatialASM, a framework to internalize active spatial mentalization into large vision-language models through training-time spatial simulation. By employing an allocentric novel-view generator, SpatialASM produces explicit trajectories that teach models when and how to adaptively invoke view planning and imagination during spatial reasoning.
[arXiv](https://arxiv.org/abs/2610.13832v1)

**Semi-OPD: Offline student rollouts**
The authors show that distilling from offline rollouts generated by the initial student (Semi-OPD) often outperforms on-policy distillation in both accuracy and training efficiency. They find that on-policy sampling is beneficial only when the initial teacher and student are highly aligned, quantified by an output-token overlap ratio.
[HF Daily Papers](https://huggingface.co/papers/2610.11291)

**MC-EvoBench: Auditing self-improvement**
The authors construct MC-EvoBench, a controlled Minecraft benchmark with 19 registered evolution chains to evaluate recursive self-improvement in agents. The benchmark separates capability change from carried-over resources and early execution failures, providing a rigorous test for whether agents truly acquire reusable capabilities from interaction history.
[arXiv](https://arxiv.org/abs/2610.13829v1)

**TabMF: Tabular foundation models for dataset representation**
The authors introduce TabMF, a framework that repurposes pretrained tabular foundation models as encoders of supervised datasets. By equipping a predictive TFM with an architecture-aware dataset-level readout and adapting it end-to-end, TabMF yields competitive representations for higher-level tasks like dataset retrieval.
[arXiv](https://arxiv.org/abs/2610.13818v1)

**Dex-One2Many: Dexterous manipulation from one demo**
The authors present Dex-One2Many, a real-to-sim-to-real framework that learns a generalizable dexterous manipulation policy from a single human video. The method abstracts the video into sequential scene graphs that guide reinforcement learning, enabling efficient exploration while preserving broad generalizability.
[HF Daily Papers](https://huggingface.co/papers/2610.12470)

**Dream to Remember: Continual steering with world models**
The authors propose Dream to Remember (D2R), which continually steers a frozen generalist policy toward new tasks while retaining earlier skills. A frozen pretrained world model rehearses prior tasks through imagined rollouts, anchoring the steering module to its earlier behavior without requiring a replay buffer or new data.
[arXiv](https://arxiv.org/abs/2610.13816v1)

## Also this week

**Reflex: Visually Grounded Reactive Humanoid Control** — A three-stage framework for learning whole-body reactive control in humanoids, focusing on catching thrown boxes. [HF Daily Papers](https://huggingface.co/papers/2610.12772)

**SoftwareGym** — An automated pipeline for synthesizing diverse GUI software environments for training computer-use agents. [HF Daily Papers](https://huggingface.co/papers/2610.12611)

**Surfing the Activation Wave** — An analytical framework for selecting hidden states in speculative decoding based on informativeness and diversity. [HF Daily Papers](https://huggingface.co/papers/2610.12939)

**On the Design Space of Continuous Diffusion Language Models** — An analysis of self-conditioning and embedding geometry in training continuous diffusion language models. [arXiv](https://arxiv.org/abs/2610.13834v1)

**When Do Biological Reasoning Models Use Their Biological Inputs?** — A mechanistic analysis showing some biological reasoning models rely more on text than foundation model representations. [HF Daily Papers](https://huggingface.co/papers/2610.00898)

**Prompteina: Natural language protein design** — The authors develop Prompteina, the first all-atom structure generation model that supports design objectives specified in natural language,. [arXiv](https://arxiv.org/abs/2610.13917v1)

**RATED: Reliability-aware tool retrieval** — The authors propose RATED, a retrieval framework that combines robust semantic matching with a lightweight contextual bandit module for dynamic. [arXiv](https://arxiv.org/abs/2610.13889v1)

**TouchScale: Visual-tactile dataset** — The authors introduce TouchScale, a 500-hour dataset of contact-rich human interaction recorded with a unified wearable setup, temporally aligning. [HF Daily Papers](https://huggingface.co/papers/2610.10288)

**Returning to Development States: When Full Sandbox Snapshots Help Coding Agents** — The authors implement Go-Explore-style search over full Daytona snapshots to improve failure recovery for coding agents. They find that full-snapshot search is most effective when long-horizon work spans multiple services with state that cannot be reconstructed from a single diff. [arXiv](https://arxiv.org/abs/2610.13841v1)
**Specialists learn faster** — The authors show that models trained on chronologically ordered knowledge ("specialists") require far fewer update steps to acquire new knowledge. [arXiv](https://arxiv.org/abs/2610.13835v1)

**ENCORE: Few-shot agentic manipulation** — The authors introduce ENCORE, which gives a coding agent a few demonstrations as evidence to read rather than as training data for robot manipulation. [HF Daily Papers](https://huggingface.co/papers/2609.37359)

**Learned abstention policies** — The authors demonstrate that a learned answer selector improves threshold-averaged utility over standard calibrated confidence filters for factual. [arXiv](https://arxiv.org/abs/2610.13821v1)

**ThinkDrag: Semantic drag-based editing** — The authors introduce ThinkDrag, a unified multimodal framework for drag-based image editing that incorporates visual reasoning. [HF Daily Papers](https://huggingface.co/papers/2610.13316)

**Work2Skill** — A framework for capturing human GUI experience for reusable agent skill induction, including a new dataset. [HF Daily Papers](https://huggingface.co/papers/2610.12984)

**RSI-Anything** — A domain-general framework for recursive self-improvement using predictive world models. [arXiv](https://arxiv.org/abs/2610.13912v1)

**X-AgentRed** — A self-evolving red-teaming framework for discovering harmful tool call paths in multi-service agent ecosystems. [arXiv](https://arxiv.org/abs/2610.13886v1)

**500B tokens later: Letting AI agents decompile a first-person shooter** — A demonstration of AI agents decompiling a first-person shooter. [source](https://momo5502.com/posts/2026-10-09-game-decompilation/)
**AI agents in a stand-up call** — An exploration of AI agents in collaborative human workflows. [source](https://atoll92.github.io/agent-standup/)

**Open-source version of Adobe Creative Suite, built with Rust, using AI** — A project recreating professional software suites using AI and Rust. [source](https://petapixel.com/2026/10/07/someone-rebuilt-free-open-source-versions-of-photoshop-premiere-and-lightroom-using-ai/)
**The gap in shared understanding of LLM capability is widening** — This is a discussion on the widening gap in shared understanding of LLM capability. [source](https://twitter.com/karpathy/status/2109361546505966046)