Also Worth Noting - 2026-08-27

Also Worth Noting

02 [Eval] SWE Refactor Bench: Can Coding Agents Complete a Long-Horizon, Whole-Repository Stack Migration? Coding agents can pass migration tests without performing a migration. The loophole is structural: existing benchmarks check behavioral correctness, so an agent that copies the original implementation and makes tests pass scores well while leaving the codebase untouched. SWE Refactor Bench closes this with 20 whole-repository migration tasks verified structurally, not just behaviorally. Teams evaluating coding agents on real-world maintenance work should treat any benchmark without structural verification as measuring test-passing, not migration. link

03 [Agent] The Handoff Tax: Continuing Non-Native Trajectories in LLM Agents Switching to a stronger model mid-task does not simply add the stronger model's capability. The receiving model pays a measurable accuracy penalty for continuing a trajectory it did not start, beyond what either model would incur on a fresh run. The amount of inherited context changes the outcome, but no configuration fully recovers the performance of a same-model run. Cost-quality routing strategies that treat model escalation as free at inference time are underestimating the real cost. link

04 [RAG] RetrievalRouter: Joint Modality and Architecture Selection for Document Retrieval Running ColPali or late-interaction retrieval on every query is accurate but expensive; dense retrieval is fast but misses evidence in complex documents. RetrievalRouter selects retrieval modality and architecture per query at inference time, breaking the fixed-pipeline trade-off between latency and recall. The routing decision is made jointly across modality and architecture, not sequentially. For teams serving retrieval in finance, healthcare, or law, this is a practical path to matching expensive pipeline accuracy at a fraction of the per-query cost. link

05 [Training] V-Rubrics: Visual Faithfulness via Rubric-Based Reinforcement Learning Scalar outcome rewards tell a vision-language model whether an answer is acceptable, not which visual inference step failed. V-Rubrics decomposes reference responses into atomic rubric items and assigns credit at that granularity, targeting the specific chart values, object references, and intermediate reasoning steps that scalar RL cannot distinguish. Hallucination on chart and object grounding tasks drops measurably compared to scalar-reward baselines. Teams doing multimodal post-training with RLHF-style pipelines should treat credit assignment granularity as a first-class design decision. link

06 [Inference] VoiceMem: Streaming Dual-Brain Memory for Real-Time Interaction Separating informational and emotional memory into parallel streams lets a duplex speech LM update both without blocking audio output. VoiceMem's left brain handles factual recall; the right brain tracks emotional state; both write and read via streaming I/O so memory retrieval does not stall generation. Under top-5 retrieval, the factual stream outperforms classical retrieval baselines. For teams building long-running voice agents, the architecture offers a concrete alternative to monolithic context windows or blocking memory lookups. link