• KV Compression Windows Create Periodic Retrieval Weaknesses: The same information can vary by up to 40 percentage points in accuracy across phases. Deployments need phase sweeps, not just average scores.

• Writable Context Shifts Information Control to the Model: Context Language Models raise BrowseComp-Plus accuracy by 11.4% while cutting compute by 21.5%.

• Follow-Up Questions Split Into Breadth, Depth, and Stopping: With the same retrieval budget, Q&D beats prompting and outperforms a 15-times-larger prompted model on two of three multi-hop QA benchmarks.

Also Notable

• The Best EngiScore Is Just 44.3 Across 1,301 Professional Engineering Tasks — Closed-loop success across multiple software tools falls to 3.6%. General computer-use scores have not translated into industrial delivery.
• Global Reading Before Targeted Zooming Raises RULER From 57.5 to 87.4 — The two methods have similar input compression rates at 3.0× and 2.9×.
• A Chinese-Specific Encoder Makes Candidate Decisions in 15 Milliseconds — It reaches 92% of Jev's average accuracy in specialized domains while running 17× faster.
• Explaining 5% of Positions Preserves Nearly All Threat-Explanation Success — This held on three of four datasets. Simple conversation structure often selects audit targets better than internal model signals.
• FRAC Approximates Power-Law Memory With Finite Exponential Modes — Log-spaced modes help state-space models avoid purely exponential forgetting while retaining bounded-state decoding.
• Agent-Designed Libraries Fail on Rigid Interfaces, Not Missing Features — Downstream agents often reimplement existing capabilities because reuse costs are too high.
• Alignment Audits Can Move Before Training — They can predict whether specific fine-tuning will amplify misaligned behavior. Data filtering usually helps on multiple-choice evaluations, but gains in open-ended dialogue remain unclear.
• Text Recognition and Glyph Accuracy Fail Independently in Video — Local text edits often damage unspecified regions. Static OCR metrics miss these temporal and editing failures.
• High-Disagreement Tokens Calibrate Confidence Better Than Selected-Token Probabilities — Average expected calibration error falls to 13.0% in white-box evaluation, versus 32.7% to 42.4% for standard full-sequence methods.
• Not Every Reflective Correction Is Worth Learning — AdviSD compares advisor scores for the same existing response with and without advice. It selects targets by the score gap, without extra executor sampling.

Read the full edition →