• One AI Tutor Matched Expert GRE Learning Gains at Far Lower Cost. It spent about $0.0052 per percentage point gained, versus $4.81 for human tutoring. The summary reported no long-term follow-up.

• JitMem Turns Long-Term Memory Credit Assignment Into Immediate Feedback. It retains raw trajectories, curates them when new tasks arrive, and beats the strongest baselines by 3.9 to 16.3 percentage points.

• Whisper Remains Usable After Losing Six Layers, but WER Rises. Label-free speech distillation cut average WER from 21.9% to 20.1%, versus 18.2% for the original model. The summary included no end-to-end speed or latency measurements.

Also Notable

• Long-Video Generation Now Has a Memory Problem. The bottleneck is shifting from frame quality to preserving identity, state, and causal changes beyond a limited context window. This survey organizes autoregressive video memory across form, function, operation, learning, and evaluation.
• Hunyuan-A13B Balances Capability, Throughput, and Deployment Cost. The open MoE model has 80B total parameters, activates 13B, and supports fast and slow reasoning modes. It was pretrained on 20 trillion tokens and evaluated on math, science, coding, and agent tasks.
• Edit Corrupted Agent State Instead of Predicting Tools. AEWM identifies and rewrites task states polluted by false assumptions and stale plans. EditAct beats the strongest baselines by an average of 3.2 to 6.7 points across six benchmarks and three agent backbones.
• WhatWorkedBench Tests Whether Research Agents Understand Experiments. Agents must predict full configuration response surfaces under limited experiment budgets. The benchmark separates running experiments from understanding changes across 36 tasks, 30 data sources, and eight workflow types.
• Strong Cross-Modal Signals Can Repair Weak Video Attention. RecCAR uses video-to-audio and video-to-motion correspondence to recalibrate weaker attention mappings in the reverse direction. It raises the human anatomy score from 0.69 to 0.75 and lowers the audio-video desynchronization metric from 0.804 to 0.752.
• VHD-Play Generates Verifiable Stateful Tool Environments. It solves a mathematical mechanism before generating each environment, so changing task rules and acceptance signals share one verifiable source. The pipeline created 3,300 environments for a few cents each and raised the average agent score across five diagnostic categories from 0.204 to 0.815.
• FLEET Makes Later Samples Remember Earlier Attempts. It reduces semantic repetition by retaining previous generations and evaluations. In the reported setting, it matched the repeated-sampling baseline's accuracy with a threefold speedup and raised LiveCodeBench Pass@32 from 59.9% to 66.2% at equal budget.
• Linear Representation Claims Need Explicit Equivalence Rules. Without them, probes, metrics, and interventions may test different propositions. The paper formalizes representational objects, generation processes, target properties, and architecture-imposed equivalences through group actions.
• Spatial Validation Shrinks Foundation-Model Gains. In old-growth forest detection, geographic foundation models lose much of their advantage over Sentinel features after spatially buffered validation. With a 10-kilometer train-test buffer, TESSERA's PR-AUC lead over Sentinel-1/2 falls to 0.04, with a confidence interval consistent with no difference.

Read the full edition →