RI

#1Reinforced Generation of Combinatorial Structures: Ramsey Numbers

MUST-READ

2026-03-11 | Google DeepMind, Google, University of California, Berkeley | 2603.09172

M=8 P=3 I=9

Nagda et al. (DeepMind) apply the AlphaEvolve framework to discover novel stochastic search algorithms that improve lower bounds for five classical Ramsey numbers and match SoTA on 23 others. The results are mathematically verified and represent genuine SoTA advances in extremal combinatorics, proving the framework's capability to generate highly specialized, non-trivial heuristics. The single most useful takeaway for us is their meta-algorithm's scoring function: instead of only rewarding valid states, they evaluate a larger, infeasible 'prospect' state and provide a dense, continuous reward based on its violation count relative to a random baseline. We should immediately steal this 'prospect evaluation' trick for AlgoEvo to smooth the reward landscape when evolving heuristics for highly constrained OR problems like VRP, where finding strictly feasible intermediate solutions is a bottleneck.

#2Advancing Automated Algorithm Design via Evolutionary Stagewise Design with LLMs

MUST-READ

2026-03-09 | Nanjing University, Huawei Noah’s Ark Lab | 2603.07970

M=9 P=8 I=9

EvoStage enhances LLM-based automated algorithm design by decomposing the generation process into sequential stages, using a multi-agent system (coordinator and coders) to iteratively refine code based on real-time intermediate execution feedback. The results are highly credible and backed by strong empirical numbers; it achieves state-of-the-art HPWL on 16 chip placement benchmarks and beats AlphaEvolve/EoH on Bayesian Optimization tasks using an incredibly small budget of just 9 to 25 evaluations. The single most useful takeaway is the shift from black-box end-to-end evaluation to stagewise intermediate feedback, where a coordinator agent reflects on mid-execution metrics to guide the next stage of heuristic design. This matters immensely for our AlgoEvo and MASPRM projects; we should immediately test pausing our VRP/scheduling environments mid-execution to feed intermediate state metrics to a coordinator LLM, which could drastically reduce the number of LLM samples we need to find optimal heuristics.

#3Code-Space Response Oracles: Generating Interpretable Multi-Agent Policies with Large Language Models

MUST-READ

2026-03-10 | Google DeepMind | 2603.10098

M=8 P=7 I=8

This paper replaces the deep RL oracle in Policy-Space Response Oracles (PSRO) with an LLM that generates interpretable Python code policies, using AlphaEvolve to iteratively refine the code against opponent meta-strategies. The results are backed by solid empirical metrics, showing that the AlphaEvolve variant achieves competitive exploitability and higher population returns than RL baselines (IMPALA) and CFR+ on Repeated Rock-Paper-Scissors and Leduc Poker. The single most useful takeaway for us is their 'context abstraction' technique—using an LLM to summarize opponent code into natural language to bypass context window limits during evolutionary search. This is highly relevant for our AlgoEvo project; we should immediately discuss implementing their two-level loop (outer meta-game equilibrium, inner AlphaEvolve refinement) and context abstraction for our multi-agent evolutionary search.

#4Autonomous Algorithm Discovery for Ptychography via Evolutionary LLM Reasoning

DISCUSS

2026-03-05 | Argonne National Laboratory, Rice University | 2603.05696

M=7 P=3 I=8

This paper applies LLM-guided evolutionary search (similar to FunSearch/AlphaEvolve) to discover novel regularization algorithms for ptychographic image reconstruction. The results are backed by solid empirical metrics, showing up to +0.26 SSIM improvements over unregularized baselines across multiple datasets. The most useful takeaway for us is twofold: first, their 'semantically-guided crossover' explicitly prompts the LLM to analyze two successful parent algorithms and intentionally merge their complementary mathematical strengths, rather than blindly recombining code. Second, the LLM autonomously discovered the benefit of embedding stateful optimizers (like Adam) and iterative sub-loops directly inside a single heuristic step. This matters for our AlgoEvo and VRP work because we should immediately steal this semantic crossover prompting strategy and ensure our evaluation API permits the LLM to generate stateful, multi-pass operators rather than just stateless, single-pass functions.

#1SLO-Aware Compute Resource Allocation for Prefill-Decode Disaggregated LLM Inference

MUST-READ

2026-03-05 | Kingsoft Cloud | 2603.04716

M=4 P=9 I=7

This paper proposes a hybrid resource allocation method for disaggregated Prefill/Decode (P/D) inference, using M/M/1 queuing theory to model prefill throughput under TTFT constraints and empirical profiling for decode. The results are real and validated on NVIDIA H200 clusters running DeepSeek-V3.1. The key takeaway for us is the validated analytical relationship between TTFT, input length, and effective prefill throughput ($TP{eff} = TP{max} - \frac{L{in}}{TTFT - T{overhead}}$). We can steal this equation to serve as a cheap, differentiable constraint in our 'GPUSched' OR formulations or fitness functions, replacing expensive simulations.

#2Serving Compound Inference Systems on Datacenter GPUs

MUST-READ

2026-03-09 | University of Illinois Urbana-Champaign | 2603.08797

M=7 P=8 I=7

JIGSAWSERVE uses a Mixed Integer Linear Programming (MILP) formulation to jointly optimize model variant selection (accuracy scaling) and fine-grained GPU spatial partitioning (MIG/MPS) for serving compound inference DAGs. The results are strongly backed by empirical numbers on real hardware (H100s), demonstrating an 11.3x capacity improvement over the closest prior work (Loki) while maintaining under 0.6% SLO violations. The single most useful takeaway for us is their specific MILP formulation, which successfully linearizes the complexities of task-graph multiplicative factors and spatial GPU segments into a tractable objective for latency and accuracy SLOs. This is highly relevant for our GPUSched project; we should extract their MILP constraints for spatial partitioning and task-graph budgeting to adapt for our own LLM inference scheduling and multi-agent resource allocation models.

#3PromptTuner: SLO-Aware Elastic System for LLM Prompt Tuning

DISCUSS

2026-03-05 | Nanyang Technological University, Unaffiliated | 2603.05087

M=5 P=8 I=7

PromptTuner is a cluster management system for LLM prompt tuning that combines a 'Prompt Bank' (retrieving similar past prompts to speed up convergence) with a hierarchical scheduler (warm/cold GPU pools) to meet latency SLOs. The authors demonstrate real-world efficacy on 32-96 GPU clusters, showing 4-8x reductions in SLO violations compared to INFless and ElasticFlow. The key takeaway for us is the 'Prompt Bank' mechanism: using K-medoids clustering on activation features to retrieve high-quality initial prompts. We should steal this initialization strategy for AlgoEvo to reduce the number of generations needed for convergence, and use the scheduling logic as a baseline for our GPUSched project.

#4BandPO: Bridging Trust Regions and Ratio Clipping via Probability-Aware Bounds for LLM Reinforcement Learning

MUST-READ

2026-03-05 | Fudan University, Shanghai Innovation Institute | 2603.04918

M=8 P=7 I=7

BandPO replaces the standard static clipping in PPO/GRPO with dynamic bounds derived from projecting f-divergence trust regions, specifically addressing a bottleneck where allowable updates vanish for low-probability tokens. Empirical results are rigorous, showing consistent gains (2-10%) on math benchmarks and, crucially, maintaining policy entropy where baselines collapse. The key takeaway is that standard clipping scales update margins linearly with probability, effectively freezing rare tokens; BandPO decouples this, allowing the model to actually reinforce novel, high-advantage tail strategies. We should implement the closed-form TV or Chi-squared variants immediately in our RL optimizers to improve exploration efficiency.