AI news digest — September 11, 2026
6 items, each with its source.
Negative Self-Distillation improves reasoning by penalizing model mistakes instead of mimicking solutions
Researchers introduced Negative Self-Distillation (NSD), a training framework that optimizes language models by diverging from self-generated flawed reasoning paths instead of imitating privileged teacher traces. The method incorporates a dynamic gating mechanism that isolates reasoning-critical tokens from general linguistic tokens during gradient updates to preserve language fundamentals.
Why it matters. Reasoning models avoid premature certainty collapse and preserve exploratory error correction during multi-step problem solving without requiring labeled solution data.
arxiv.orgKashin-DCT enables stable two-bit language model quantization with low computation overhead
A new post-training quantization method adapts the Kashin decomposition using sign-randomized Discrete Cosine Transforms, lowering per-iteration transformation complexity from quadratic to log-linear time. The approach factors weight matrices into bounded components suitable for native two-bit clustering while preventing the numerical divergence issues common in second-order quantizers.
Why it matters. Native two-bit execution becomes practical for large transformer layers without suffering numerical breakdown during sequential error compensation.
arxiv.orgContinuous adaptive denoising enables robust humanoid locomotion under corrupted depth vision
Researchers developed CAP, a unified single-stage locomotion policy for humanoid robots that couples a learned depth-denoising world-model encoder with a proprioceptive variational encoder. Instead of abruptly switching between perceptive and blind control regimes, the policy continuously adapts across varying degrees of visual corruption and was verified on the Unitree G1 robot.
Why it matters. Humanoid robots can traverse complex outdoor terrain without tripping when onboard depth cameras suffer intermittent lens occlusion or lighting artifacts.
arxiv.orgReal-time Jacobian estimation achieves sub-millimeter dexterous in-hand robotic pen writing
An embodied control framework enables anthropomorphic robotic hands to manipulate and write with a pen using online task Jacobian estimation on physical hardware. Operating entirely on a laptop CPU without prior simulation training or demonstration datasets, the system adapts within 18 seconds to achieve an average writing precision of 0.6 millimeters.
Why it matters. High-precision in-hand manipulation becomes deployable on physical robots without requiring millions of reinforcement learning simulation steps or teleoperated demonstration libraries.
arxiv.orgMindTopo benchmark reveals multimodal foundation models struggle with topological spatial reasoning
The MindTopo benchmark evaluates 14 multimodal foundation models across 11,030 procedural tasks assessing topological invariants such as continuity, separation, order, enclosure, and knots. Across both static visual reasoning and closed-loop agent planning, all evaluated foundation and video generative models fell significantly behind human baselines.
Why it matters. Vision-language models remain inadequate for physical assembly and routing tasks that depend on topological spatial invariance rather than metric coordinate estimation.
arxiv.orgRAG-Safety-Bench isolates how retrieval-augmented generation degrades safety guardrails in language models
A new evaluation framework measures how retrieval-augmented generation impacts safety alignment across five open-source language models across four controlled retrieval conditions. The empirical results show an inverse relationship between benign capability and safety compliance, demonstrating that even safe retrieved context can trigger unsafe completions to adversarial prompts.
Why it matters. Base model safety alignment guarantees fail to hold once retrieval pipelines are attached, requiring application-level safety filtering on retrieved inputs.
arxiv.orgFeed, daily deep-dive and bytes — readable offline, with push alerts for the topics you follow.