Tag: #llm90 results
This survey frames multi-agent communication via the Five Ws, tracing evolution from MARL's hand-designed protocols to emergent language and LLM-based systems. It highlights trade-offs in interpretability, scalability, and generalization across paradigms. Practical design patterns and open challenges are distilled for hybrid systems.
Introduces BLPO to optimize prompts for multimodal LLM-as-a-judge evaluating AI images. Overcomes context limits by converting images to text representations. Outperforms baselines on four datasets with three LLM judges.
Nvidia's DMS compresses LLM KV cache up to 8x, reducing memory costs without accuracy loss. Enables longer chain-of-thought reasoning and more parallel paths. Outperforms heuristic eviction and paging methods.
MiniMax released open-source M2.5 and Lightning models, rivaling top models at 1/20th Claude Opus cost. MoE architecture activates 10B of 230B params; excels in agentic tasks like Office files. Used internally for 80% code commits.
MiniMax releases open-source M2.5 and Lightning models, matching state-of-the-art at 95% lower cost via API. MoE activates 10B of 230B params; excels in agentic tasks like Office files. Used internally for 80% code commits.
Google appears set to release Gemini 3.1 Pro soon, with model references already spotted in related arenas. This follows recent launches like Zhipu's open-source GLM-5 and DeepSeek's upgraded model with larger context window.
Locomo-Plus benchmarks cognitive memory in LLM agents under cue-trigger disconnects, focusing on latent conversational constraints. It proposes constraint consistency evaluation over string-matching. Reveals gaps in existing memory systems.
Apple researchers demonstrate that reasoning trace length serves as a simple, effective confidence estimator in large reasoning models. It performs comparably to verbalized confidence across models, datasets, and prompts, acting complementarily. The work shows reasoning post-training alters the trace-confidence relationship.
Three new Chinese AI models launched for OpenClaw: DeepSeek V4, GLM-5, and Qwen3-Coder-Next. They deliver frontier-level agentic capabilities. Pricing is a fraction of Western models.