New studies focus on catching silent agent failures, rewarding each reasoning step in RAG, and training models to abstain until evidence is sufficient. Benchmarks also show that larger models do not always win on accuracy or efficiency.
A test on 25 Nutrition5k meals compares seven multimodal LLMs on calorie estimation under a 20% error threshold. Open-weights Muse Spark 1.3 hits 48%, while Qwen 3.8 27B stops at 16%. The ranking doesn't track model size—a useful signal for anyone ev...
A benchmark on 14 cross-system tasks and three MCP servers shows that TrueForge with Opus 4.8 solves 11/14 tasks like Claude Managed Agents, but uses 63% fewer tokens and costs 30% less per run. With GLM-5.2 the cost drops to $3. Native tracing, sand...
A training framework teaches grounded QA models to answer only when evidence becomes sufficient, locating the exact point where abstention gives way to a response. On HotpotQA, 2WikiMultiHopQA and MuSiQue, with Qwen2.5-3B-Instruct and LoRA, the metho...
PRO-Step introduces step-level supervision for Retrieval-Augmented Generation. Instead of evaluating only the final answer, a generative process reward model checks logical validity and evidential grounding at every step of multi-hop reasoning. The m...
A study across 400 trajectories shows outcome-only LLM judges catch 84% of loud faults but only 45% of silent ones while flagging 33% of correct trajectories. A step-rubric judge reaches 77% silent recall with zero false alarms at 3x cost. No judge r...
