Search: “efficiency”

7 stories in the last 24h

LLM-Falsifier uses large language models with semantic prompting to minimize STL robustness and find counterexamples in cyber-physical systems more efficiently.

The paper formulates falsification of Signal Temporal Logic specifications as robustness minimization and leverages iteratively prompted LLMs as optimizers. It exposes the LLM to semantic information absent from numerical optimizers, including natural-language signal names, output trajectories, and critical-time witnesses for the minimum robustness value. On ARCH-COMP falsification benchmarks, LLM-Falsifier required fewer simulations than surrogate-based, Bayesian, and search-based tools on 14 of 21 specifications.

arXiv cs.AI / cs.LG / cs.CLSep 17, 17:40 UTC · 20h agoAI research

dQwen3.5: Hybrid-Attention Diffusion Language Models

Researchers adapt Qwen3.5 hybrid-attention backbones (0.8B-9B) into diffusion language models, reaching training loss in half the tokens versus full attention.

Researchers converted pretrained Qwen3.5 models at 0.8B, 2B, 4B, and 9B scales into the dQwen3.5 family of diffusion language models. The work addresses hybrid attention-RNN backbones, whose RNN layers are structurally causal and hard to bidirectionalize. Against a full-attention control, the hybrid backbone reaches a given training loss in roughly half the tokens. The resulting models match full-attention DLMs in any-order decoding and perform strongly under parallel decoding.

arXiv cs.AI / cs.LG / cs.CLSep 17, 17:39 UTC · 20h agoAI research

Workspace Models: Lightweight Robotic Memory via Saliency-Driven Supervision

Researchers propose workspace tokens, a lightweight latent robot memory distilled from train-time VLM saliency queries, improving policy performance without in-loop VLM calls.

The paper introduces workspace tokens, a latent memory representation for robotic manipulation policies trained by using a VLM to identify task-relevant current and historical information and distilling it via a set-reconstruction decoder loss. At deployment the workspace token serves as a drop-in replacement for observations, letting policies solve memory-intensive tasks without expensive in-the-loop VLM queries. Experiments in simulation and on hardware show it is both more lightweight and yields better policy performance than conditioning on full histories or in-loop VLM compression.

arXiv cs.AI / cs.LG / cs.CLSep 17, 17:59 UTC · 20h agoAI research

Score Centering Stabilizes Off-policy Reinforcement Learning

Paper attributes RL instability under training-inference mismatch to accumulating drift and introduces score centering, an additive correction that stabilizes training.

The authors show that instability of LLM reinforcement learning under the training-inference mismatch (TIM) is primarily caused by drift, a persistent bias between training and inference engines that accumulates each step. They derive an additive score centering correction that cancels drift, and on models from 0.6B to 30B parameters it matches or outperforms importance-sampling methods under quantization as mismatch grows. Because it is additive, score centering composes with importance sampling, outperforming pure importance-sampling baselines in staleness experiments.

arXiv cs.AI / cs.LG / cs.CLSep 17, 17:58 UTC · 20h agoAI research

An Empirical Study of Harness Design for Coding Agents

Empirical study of coding agent harnesses across 176 settings shows context management prevents overflow failures and bash-only interfaces cut costs for capable models.

The study uses a lightweight coding harness with a fixed execution loop while varying planning, action space, and context management, evaluating 176 matched settings across four models on SWE-Bench Verified and Terminal-Bench 2.1. Context management becomes more valuable as context-window budgets tighten, mostly by preventing context-overflow failures, with rule-based elision before LLM summarization the strongest strategy. Planning shifts from an accuracy scaffold for weaker models to a cost saver for stronger ones, and bash-capable models work well with a bash-only interface at substantially lower cost, informing model- and budget-aware harness design.

arXiv cs.AI / cs.LG / cs.CLSep 17, 17:58 UTC · 20h agoAI research 3 sources1

Agile-WAM: An Agile Tactile World Action Model for Contact-Rich Robot Control

Researchers introduce Agile-WAM, a tactile world action model using direct vision-tactile-to-action flow matching for agile contact-rich robot control.

Agile-WAM encodes visual and tactile observations into a shared latent and jointly generates action chunks plus future visual and tactile latents via flow matching, avoiding large pretrained generative backbones. Multi-horizon multimodal prediction supervises visual latents at longer offsets while capturing abrupt tactile contact dynamics in the next frame. Across nine simulated and five real-world manipulation tasks it achieved a 29.4% relative success-rate gain over the strongest baseline with 11.9 ms inference latency.

arXiv cs.AI / cs.LG / cs.CLSep 17, 17:43 UTC · 20h agoAI research

On-Demand Attention: Language Models Know When to Recall

On-Demand Attention uses a lightweight recall head to selectively invoke global attention, cutting long-context decoding costs in vLLM with minimal quality loss.

Researchers introduce On-Demand Attention (ODA), a local-first decoding method in which a lightweight recall head predicts when global attention benefits the next token. Only the recall head is trained, leaving pretrained weights and the complete KV cache unchanged. GPU-side conditional execution implemented in vLLM converts reduced global reads into practical decoding speedups at long context lengths. Experiments across Qwen and Gemma models, including hybrid-attention backbones, show selective recall recovers most of the performance lost under local attention.

arXiv cs.AI / cs.LG / cs.CLSep 17, 17:24 UTC · 20h agoAI research