Large Language Models
1
Subquadratic Attention with a Hierarchical Tree-Structured Indexer
Zyphra Hierarchical Attention builds a tree of recursively compressed key summaries over the KV cache; queries descend the tree to pick a sparse key set, so total cost including routing grows subquadratically. It can be added to pretrained models via a short KL distillation over a few billion tokens while preserving long-context retrieval, reasoning, math and agentic coding, and is robust to trees up to 6 levels deep. The authors project it beats competing methods beyond 1M tokens.
Why it mattersUnlike sparse attention whose indexer remains quadratic, ZHA's routing itself scales subquadratically, and it can be retrofitted onto existing models cheaply. If the performance model holds, it is a credible path toward 100M+ token contexts for agents.
arXiv:2610.42112 ↗PDF ↓
2
Latent Sparse Attention: Towards Subquadratic Ultra-Long Context Attention
LSA hierarchically clusters keys in a learned latent space so queries score clusters rather than every key. D-LSA retrofits pretrained models via distillation and reports large prefill and decode speedups over full and sparse attention at 10M tokens with similar accuracy; N-LSA, trained on next-token loss alone, matches full attention with a 1.6x per-layer training speedup at 128k tokens.
Why it mattersAppearing the same day as Zyphra's ZHA, it shows hierarchical key routing is converging as the leading recipe for 1M–10M token attention. Offering both retrofit and native-training variants makes it relevant to labs at different stages.
arXiv:2610.42147 ↗PDF ↓
3
CRoS: Causality Reward on Scrambled CoTs Improves Accuracy and Reasoning Faithfulness
Since RLVR rewards only correct answers, chains of thought may not causally drive the result, and prior causality metrics that delete or corrupt steps change content too. CRoS instead scrambles the order of reasoning steps. On Reasoning Gym with Gemma-3-4B it reaches 46.8% on five unseen logic/game datasets versus 43.8% for the best baseline, and raises hint acknowledgment (a faithfulness measure) from 20.2% to 34.0%.
Why it mattersIt offers a cheap, content-preserving way to reward reasoning that actually drives the answer, improving both accuracy and faithfulness. Faithful CoT matters for monitoring and safety as RL-trained reasoners spread.
arXiv:2610.42054 ↗PDF ↓Project page ↗
4
On the Origin and Evolution of Training-Inference Mismatch in LLM Reinforcement Learning
The sampler and trainer in on-policy LLM RL disagree, making data effectively off-policy. The paper splits the error into a roughly constant numerical discrepancy and a self-reinforcing policy lag from asynchronous sampling; importance weighting fails on Qwen3-1.7B once effective sample size drops below one half. The proposed DLC centers the discrepancy, clips the lag ratio and resyncs the sampler, keeping KL from on-policy training within twice the run-to-run noise and preventing collapses.
Why it mattersTraining-inference mismatch is a known cause of unstable large-scale RL runs, and this gives a principled diagnosis plus a practical correction. The ESS-below-one-half resync rule is easy for RL infrastructure teams to adopt.
arXiv:2610.40945 ↗PDF ↓
AI Agents
5
Mage-VL-Cowork: A Full Training Recipe to Learn Hybrid Computer Use From Evolved Harnesses
Mage-VL-Cowork is a 4B agent that combines GUI interaction with code and terminal execution in a single policy, trained through a five-stage curriculum of midtraining, SFT and RL. Its Mage-CUA-Forge pipeline synthesizes 293,806 desktop and web tasks across 6,915 skills at 82.6% lower recurring cost, and evolved teacher harnesses supply stage-specific supervision. It reaches 66.0% on OSWorld-Verified, which the authors call state of the art for its size.
Why it mattersA 4B model at 66% on OSWorld-Verified suggests capable computer-use agents can run locally and cheaply. The full recipe, from task synthesis to harness evolution, is a reusable blueprint for training hybrid GUI-plus-code agents.
arXiv:2610.40747 ↗PDF ↓Project page ↗
6
When Requirements Interact: Scaling Long-Horizon Coding via Interdependent Specifications
Coding agents must satisfy interacting requirements without breaking existing dependencies, which current benchmarks barely test. CodeSpecWeave uses a synthesis agent to build coupled requirements and executable tests from real developer seeds and a review agent to audit multi-model solutions, yielding 4,670 training tasks across 2,865 repos. RL on them improves results on DeepSWE-v1.1, and the released CodeSpecWeave-Bench (100 tasks, 1,044 specs) separates frontier models.
Why it mattersReal software work is about satisfying interacting constraints, not isolated bug fixes, and this supplies both RL training data and a harder benchmark for it. It comes from Xiaomi's MiMo team, which is pushing its own agentic coding models.
arXiv:2610.40256 ↗PDF ↓
7
TraceSpec: Understanding and Improving Speculative Decoding for LLM Agents
Aggregate speculative-decoding metrics hide how agents' decisions vary across steps. TraceSpec pairs trajectory-level evaluation with adaptive draft training and releases TraceSpec-Traj (10,000 train and 2,500 test states over browsing and software engineering). It finds acceptance can rise with context length overall while falling on late browsing actions, and improves replay throughput by 12.5–22.6% over its DSpark base.
Why it mattersAgent workloads dominate inference cost, and this shows speculative decoders tuned on chat-style data leave speed on the table. Trajectory-aware draft training is a practical lever for serving agents faster.
arXiv:2610.41177 ↗PDF ↓
