Bi-Daily Research Intelligence Briefing

Li et al. introduce MIPLIB-NL, a benchmark of 223 industrial-scale MILP instances (up to 10^7 variables) reverse-engineered from MIPLIB 2017, enforcing strict model-data separation. Results are sobering: SOTA models like GPT-4 and fine-tuned OR-LLMs drop from ~90% accuracy on existing toy benchmarks to ~18% here, failing primarily on structural consistency and index handling at scale. For us, the key takeaway is their "Loop-Based Structural Scaffold" taxonomy—a method to compress massive industrial formulations into compact LLM prompts via model-data separation. This is a mandatory read for our OR-Bench project, as it demonstrates that current evaluations are effectively measuring overfitting to toy problems rather than genuine modeling capability.

AMPD introduces a disaggregated serving framework tailored for multi-round LLM agents, utilizing an offline ILP solver to optimize resource allocation (TP/DP configurations) and an online adaptive routing mechanism to handle incremental prefill tasks. The results are strong, showing 67-340% improvements in SLO attainment over vLLM and NVIDIA Dynamo by dynamically routing incremental prefill to decode workers when slack exists. For our 'GPUSched' project, the key takeaway is the specific ILP formulation (Eq. 5) for partitioning prefill/decode resources under global GPU constraints, and the insight that multi-agent workflows create a unique 'incremental prefill' bottleneck that standard disaggregation handles poorly.