2026-09-25

TL;DR - SWE-Prometheus is a 60-repository benchmark that tests whether coding agents can independently identify and implement engineering-governance improvements while preserving behavior. It distinguishes merely adding governance artifacts from producing verified, execution-backed gains.

• The benchmark evaluates six governance dimensions using paired evidence, clean-environment probes, behavior gates, and two independent teacher ratings.
• Across ten models on a shared 22-repository subset, mean Normalized Governance Improvement ranged from 0.0568 to 0.5760, with behavior-breakage rates from 0% to 23%.
• A repository-blind template scored 0.272 mean NGI on a frozen ten-repository batch, but improved only Tests & CI, Quality Gates, and Documentation—not reproducibility or dependency/security governance.
• Results support reporting improvement, behavior preservation, evidence quality, and coverage together rather than relying on a single aggregate score.

TL;DR - Coding agents can synthesize reusable programs for generalized task and motion planning, outperforming traditional planners and LLM baselines across simulated robotics environments. The results suggest agent-written code is a strong, compute-efficient baseline for solving unseen planning instances with less domain-specific engineering.

• Claude Code and Codex agents generated 980 programs across 28 KinDER and PDDLStream environments, evaluated over 98,000 held-out episodes.
• Agent configurations achieved 56%–95% mean success, versus 47% for hand-engineered planners on the 16 environments where planners were available.
• Generated programs maintained higher success as object counts increased while using roughly an order of magnitude less computation per instance.
• Agent logs showed iterative simulator use to calibrate physical models, test edge cases, and refine planning strategies.

TL;DR - IterSynth is a deep-search agent framework that separates planning from evidence synthesis and maintains a compact evolving summary instead of a growing search history. This reduces role coupling and context noise, improving long-horizon search performance across multiple models.

• Alternates between a Planner that identifies information needs and a Synthesizer that integrates evidence into the persistent summary state.
• Introduces Role-Decoupled Policy Optimization, combining terminal rewards, turn-level rubric evaluations, and role-specific advantage estimates.
• IterSynth-8B averages 50.7 across five deep-search benchmarks, outperforming the strongest prior agent at or below 8B parameters by 4.2%.
• The prompting paradigm also provides substantial reported zero-shot gains over ReAct-style approaches on frontier proprietary models.

TL;DR - Synthetic Hospital is an open, fully synthetic benchmark for evaluating language models on realistic longitudinal electronic health records without exposing protected health information. Its verifiable ground truth and simulated EHR infrastructure enable reproducible testing of clinical AI systems.

• Includes 1,268 synthetic patients and 5,602 encounters grounded in public medical-education sources and standard clinical ontologies.
• Provides complete provenance plus interoperability APIs, role-based access, and a function-calling interface modeled on real hospital systems.
• Physicians distinguished synthetic from real charts at only 53% accuracy in a blinded review.
• The best of 10 evaluated models achieved 0.73 severity-weighted F1 on longitudinal problem-list reconstruction and missed roughly half of clinically relevant findings in chart summaries.

TL;DR - This paper presents evidence that Transformers can linearly superpose distinct text streams, producing a combined next-token distribution resembling the average of their individual distributions. Lightweight fine-tuning and guided decoding restore and exploit this behavior to generate two coherent continuations from one forward pass.

• The authors call this behavior the Superposition Linearity Hypothesis.
• Experiments suggest linear superposition is intrinsic to Transformer architecture but weakens during pretraining.
• Lightweight fine-tuning substantially reduces divergence between superposed predictions and averaged individual predictions.
• A guided decoding method disentangles the combined output into two simultaneous, coherent generations.

TL;DR - WanPE is a 397B-parameter prompt-enhancement model that converts user requests into shot-level cinematic plans for text-to-video generation. Paired with Wan3.0, it substantially improves human preference, especially for 30-second videos.

• Trained on 1.05M real-world videos using video-grounded reverse construction rather than forward prompt rewriting.
• Semantic-Consistency GRPO preserves user requirements across shots and over time.
• WanPEval covers 5–30-second videos and includes roughly 11K blind pairwise human assessments.
• Human preference gains over raw prompts range from 10.66–18.84 points at 5–15 seconds to 50.86 points at 30 seconds.

TL;DR - A study of 13 multimodal LLMs finds that common visual-text alignment scores can remain high even when visual tokens are corrupted and task accuracy collapses. It attributes this “alignment illusion” to shared language-model weights and proposes the principal-angle gap as a more reliable geometric diagnostic.

• CKA, SVCCA, MIR, and leading principal-angle cosine often failed to distinguish genuine visual tokens from Gaussian noise.
• Anisotropic MLP down-projections can pull visual and text representations toward shared output directions, creating largely one-dimensional, weight-induced similarity.
• The proposed principal-angle gap measures the difference between the top two principal-angle cosines to separate this effect from multidirectional visual structure.
• Under graded corruption, the gap tracked task accuracy more consistently, though irrelevant-image tests showed that internal geometry still need not correspond directly to task performance.

TL;DR - YODAS v3 is an open, weakly labeled speech corpus containing over 1.1 million hours of 48kHz multichannel audio across 147 languages. Its unprecedented scale, language coverage, and stereo fidelity could support research in multilingual speech recognition and neural audio codecs.

• Released under CC BY 3.0 and described as the largest open speech dataset to date.
• Introduces collection techniques designed to gather more language-balanced speech data.
• Includes over 10,000 hours for 22 languages and over 5,000 hours for 73 languages.
• Provides analyses of language coverage, audio quality, and transcription quality, plus baseline speech-recognition and neural-codec models.

TL;DR - KernelOPT is a multi-agent system that optimizes compiler-generated Triton GPU sub-kernels while preserving vendor library calls and validating the reassembled model end-to-end. It improves performance over torch.compile on KernelBench while retaining the compiler baseline when no safe optimization passes verification.

• Five profiling-guided LLM agents search for optimizations within the compiled model’s existing structure.
• A four-gate cascade checks static validity, multi-seed correctness, float64-fallback model behavior, and performance.
• The system supports PyTorch nn.Module models, standalone Triton kernels, and Helion kernels.
• Across 250 KernelBench problems, geometric-mean speedups over torch.compile were 1.40×, 1.15×, and 1.07× for Levels 1, 2, and 3, respectively.

TL;DR - Rolling-WAM accelerates closed-loop robotic manipulation by spreading joint video-action denoising across successive replanning cycles instead of recomputing the full prediction horizon each time. It achieves competitive performance with a 4.5× steady-state replanning speedup over standard joint World Action Models.

• Maintains a sliding window of video-action chunks at staggered noise levels.
• Fully denoises the imminent action chunk while progressively refining future chunks.
• Reuses evolving visual-action context as new camera observations advance the window.
• Evaluated on LIBERO, RoboTwin, and a real-world Unitree G1 humanoid.