Research Radar — July 25, 2026

Items 4–10

dLLM

Don't Commit Alone: Joint Token Commitment in Diffusion LLMs

Marker gate defers commitment; last-n layers re-run for joint decoding before argmax.

dLLMs commit multiple positions independently, accumulating factorization error when positions are dependent. CoCommit inserts a marker-gated coordination pass: the backbone's last-n layers re-run on the marked commit set, approximating joint-mode decoding. Reuses existing weights (one extra partial forward pass); improves accuracy on all 6 benchmarks on LLaDA2.1-mini with largest gains on reasoning tasks. arXiv July 5, 2026.

9 entries removed on 2026-09-10 as repeats of earlier reports: 2507.11097 (first covered 2026-07-07), 2606.25331 (first covered 2026-07-01), 2606.16939 (first covered 2026-07-01), 2605.03228 (first covered 2026-07-10), 2607.00481 (first covered 2026-07-04), 2606.10525 (first covered 2026-07-03), 2607.03236 (first covered 2026-07-15), 2604.23130 (first covered 2026-07-07), 2606.08292 (first covered 2026-07-10).

Notes

DIJA vs. Hao Peng DIJA (W30 #2): Two distinct ICLR 2026 papers on dLLM jailbreaks — W30 covered 2507.02983 (Hao Peng et al.); today's #1 is 2507.11097 (Zichen Wen et al., DIJA framework with quantitative ASR results across four models).
Substack-style layout: Today's report extends the figure-per-item format to all 10 entries. Items 4–10 carry thumbnail SVG figures in the HTML and direct arXiv HTML image URLs in the .md — same editorial treatment for every paper.
iLLaDA (#2) weekly nomination: First masked dLLM base model to exceed Qwen2.5-7B on average. The remaining instruct-model gap is RL alignment, not architecture — strong W31 candidate.
CircuitLasso (#3): Makes circuit discovery practical at production SAE dictionary scale (65K features); direct unblock for mech-interp research teams running large-scale SAEs.
Ablation-Reversible Heads (#10): Methodological critique — has implications for how the field evaluates mechanistic role claims, particularly for induction heads and copy-suppression heads that have been widely cited.