• Multi-Agent Debugging Moves from Vibes to Numbers. TraceElephant turns failure attribution into an explicit benchmark, with full execution traces lifting attribution accuracy 76% over agent-output-only views.

• Frozen Base Models Can Still Surface Key Evidence. HiLight trains a side Actor that adds emphasis tags to the input; the main model stays frozen, and the learned policy zero-shots to closed-source APIs.

• Hybrid Routing Becomes Something the Model Learns. RouteLMT replaces hand-tuned escalation thresholds with marginal-gain prediction read off the small model's own token representations, validated only on translation though.

• Audio Generation Catches Up on the Unified-Architecture Playbook. UniSonate stuffs TTS, TTM, and TTA into one text-instruction model, hitting SOTA on the first two but only "competitive" on TTA, the typical fault line for unification.

Also Notable

• Toward a Shared Vocabulary for "World Model". A capability-tier × scaling-laws taxonomy tries to unify the muddled definitions floating around agent research into comparable axes.
• Agent Discovery and Matching Finally Has a Benchmark. Picking the right agent from a pile to do a task used to mean directory browsing; this turns it into a measurable problem.
• Watermark-Style Decoding Constraints Cut Context-Faithfulness Hallucination. No retraining or weight modification, applied at decode time. Worth evaluating as a back-end RAG safeguard.
• KG-RAG Tackles Semantic Mismatch via Evidence Path Mining. Instead of fixing the graph structure, this approach extracts evidence paths to align with query semantics.
• Probing for "Preference Heads" Inside LLMs. Mechanistic interpretability tests whether dedicated attention heads handle personalization, tracing back what is currently done with prompts and fine-tuning.
• Cloud Visual Localization Without Sending Images or Keypoints. Geometric bilinear obfuscation replaces raw image features for pose estimation. CVPR work.
• NL2SQL Benchmark Finally Covers Ambiguity and Unanswerable Queries. Multi-source ambiguity and unanswerability are two thorny cases existing evaluations skip; this one folds them in.
• Do LLMs Reuse the Same Neural Mechanism Across Syntactic Constructions? A linguistically grounded probe of fine-grained internal mechanism.
• Gloss-Free Sign Language Translation via Selective Contrastive Learning. Aligns visual signs and text directly without costly gloss annotations to bridge the modality mismatch.

Read the full edition →