• Relevant Documents Are Not a Good Evidence Set. AdaTutoRank adjusts tutoring intensity to candidate-set quality and distills guidance gains into per-token advantages, separating key evidence from “free-rider” documents.

• MALA Scores Every Valid Attention Pair Before Allocating Compute. In 128K-context operator tests, it accelerated training forward passes by 2.2×, backward passes by 3.0×, and decoding by 1.6×.

• Extreme Visual Compression Must Correct the Student’s Own Trajectory. With only 5% of visual tokens, LT-OPD raised Qwen3.5-4B’s average retained performance across nine benchmarks from 68.6% to 82.3%.

Also Notable

• CIS Tightens Importance-Weight Caps by Token Confidence — It targets probability mismatches between RLVR training and inference engines. Across five math-reasoning benchmarks, it beat evaluated baselines on average for all three mixture-of-experts models.
• KV Heads Can Share Complete Causal Coverage — CoWA assigns different parts of long-range history to different heads. In 128K operator tests, training forward and backward latency fell to 1/7.4 and 1/8.6 of full attention.
• Meaning-Preserving Transformations Expose Asymmetric Generalization — Models largely retain task performance after adapting to a transformation, but harmful-output rates can rise sharply. Alignment appears more dependent on the input’s surface distribution.
• Solving Every Intermediate Step Does Not Ensure the Answer — Across 354 NuminaMath problems and six models, OracleLadder found composition gaps were the largest failure category for every model.
• External Memory Does Not Consistently Beat Native Memory — SCLATE compresses cross-month scenarios into hours with a shared event clock. Results vary widely across models using identical harnesses and memory configurations. Post-training Qwen3.5-4B improved without modifying either component.
• NTRL-Code Self-Improves From Noisy Unlabeled Instructions — It creates semantic anchors through conservative denoising, then builds proxy objectives by aggregating candidate programs through their abstract syntax trees.
• Models Improve Weak Starters More Consistently Than Editable Competent Baselines — OptiArena fixes five editing rounds, scaffolding, and resource budgets. It also checks whether editing degrades competent baselines.
• Delayed QA Supervision Can Train Test-Time Memory — Placing questions after many unrelated events directly trains long-term fact retention and updates. How much of the gain comes specifically from the delay still needs isolation.
• World SLAM Model Turns SLAM Into Long-Horizon Memory — It brings persistent state updates, durable memory, and back-end error correction into end-to-end navigation instead of treating SLAM as an upstream module.
• Ad-Account Detection Agents Should Investigate, Not Decide — In compromised business ad-account detection, a neuro-symbolic stage makes the final decision using symbolic rules, naive Bayes calibration, and a data-tuned contradiction layer. Precision rose from 0.250 to 0.446, while recall fell from 0.920 to 0.660.
• PULSE Selects Demonstrations by Internal Utility Features — It finds sparse-autoencoder features associated with gains for the target model, replacing retrieval based only on textual similarity.
• Ignored Curvature Couplings Cause Low-Rank Collapse — At high compression rates, SDLRT restores important singular directions with a lightweight compensation buffer. Negative feedback then stabilizes each layer’s rank.

Read the full edition →