RI

#1CausalEvolve: Towards Open-Ended Discovery with Causal Scratchpad

MUST-READ

2026-03-15 | Carnegie Mellon University, MBZUAI, Hong Kong Baptist University, The University of Sydney | 2603.14575

M=8 P=9 I=8

CausalEvolve enhances LLM evolutionary search frameworks (like AlphaEvolve and ShinkaEvolve) by introducing a causal scratchpad that extracts outcome- and procedure-level factors to explicitly guide program mutations via a Multi-Armed Bandit. The results are backed by solid empirical numbers, showing it outperforms the state-of-the-art ShinkaEvolve across four algorithmic and mathematical tasks (Hadamard, Autocorrelation, Circle Packing, AIME) by up to 9.1% in best-found scores. The single most useful takeaway for us is their 'surprise detection' module: using LLMs to perform abductive reasoning when a seemingly good combination of factors yields a score drop, thereby uncovering hidden confounders and generating new search directions. This is highly relevant to our work; we should immediately evaluate their MAB-driven causal intervention strategy and procedure-level factor extraction to improve the sample efficiency and memory mechanisms in AlgoEvo.

#2CreativeBench: Benchmarking and Enhancing Machine Creativity via Self-Evolving Challenges

MUST-READ

2026-03-12 | Tsinghua University, Peking University, Southern University of Science and Technology, University of Bristol, The Hong Kong University of Science and Technology (Guangzhou), Xi’an Jiaotong University | 2603.11863

M=8 P=6 I=9

Wang et al. introduce CreativeBench to evaluate LLM code generation creativity and propose EvoRePE, a representation engineering technique that extracts a 'creativity vector' from AlphaEvolve search trajectories to steer model activations at inference time. The results are backed by solid empirical evaluations, showing that injecting this vector improves novelty and correctness even without running the full evolutionary search. THE SINGLE MOST USEFUL TAKEAWAY: We can run our evolutionary search (e.g., AlgoEvo) offline to collect (baseheuristic, evolvedheuristic) pairs, compute the PCA of their hidden state differences, and inject this vector during standard inference to force the model into an exploratory mode. This is highly relevant to our work as it offers a completely new, training-free mechanism to solve the sample efficiency and scalability bottlenecks in LLM evolutionary search.

#3KernelFoundry: Hardware-aware evolutionary GPU kernel optimization

MUST-READ

2026-03-12 | Intel Corporation | 2603.12440

M=8 P=6 I=9

KernelFoundry is an LLM-based evolutionary framework for GPU kernel optimization that combines MAP-Elites quality-diversity search with meta-prompt co-evolution and gradient-informed mutation hints. The results are rigorously backed by numbers, showing a 2.1x speedup over the AI CUDA Engineer baseline on KernelBench L2 and successful optimization of Llama 3 operations. The single most useful takeaway for us is their 'Gradient-Informed Evolution' technique: they track parent-to-child fitness transitions in the MAP-Elites archive to compute pseudo-gradients across behavioral dimensions, which are then translated into specific natural language mutation hints for the LLM (e.g., 'positive gradient in memory -> hint: add shared memory tiling'). While we do not write low-level GPU kernels, this exact architectural improvement—alongside their meta-prompting to prevent context pollution—is highly transferable and should be immediately tested in AlgoEvo and EvoCut to improve our search signal and sample efficiency.

#4Procedural Generation of Algorithm Discovery Tasks in Machine Learning

MUST-READ

2026-03-18 | University of Oxford, University College London, University of California, Santa Barbara, University of Wisconsin–Madison, Delft University of Technology | 2603.17863

M=6 P=8 I=8

This paper introduces DiscoGen, a procedural generator that combinatorially creates millions of algorithm discovery tasks (varying domains, editable modules, and datasets) with strict meta-train/meta-test splits to evaluate and train Algorithm Discovery Agents (ADAs). The results are backed by extensive empirical evaluation of open-source LLMs on a fixed subset (DiscoBench), demonstrating that current ADAs struggle with multi-module discovery and that prompt-tuning over a diverse set of procedurally generated tasks significantly improves generalization. The single most useful takeaway is the combinatorial task generation approach (toggling which modules are editable vs. fixed), which provides a brilliant blueprint for creating an autocurriculum to train our 'evolver' agents. This matters immensely for us; we should immediately consider using DiscoGen to evaluate AlgoEvo, and adapt their procedural task generation strategy to create diverse OR/routing environments for our own RL-infused evolutionary search training.

#1MetaClaw: Just Talk -- An Agent That Meta-Learns and Evolves in the Wild

MUST-READ

2026-03-17 | Carnegie Mellon University, UC Berkeley, UNC-Chapel Hill, UC Santa Cruz | 2603.17187

M=8 P=7 I=9

MetaClaw is a continual learning framework for LLM agents that combines gradient-free skill evolution (distilling failures into reusable prompt instructions) with asynchronous RL fine-tuning guided by a process reward model. The results are backed by solid empirical gains, showing an 8.25x improvement in end-to-end task completion on a 30-day simulated CLI benchmark. The single most useful takeaway for us is their 'Skill Generation Versioning' mechanism: when co-optimizing discrete prompts/heuristics and continuous model weights, you must strictly separate support data (used to evolve the heuristic) from query data (collected after the update) and flush the RL buffer upon evolution to prevent training on stale rewards. This directly addresses the continuous learning and RL-infused evolution bottlenecks in AlgoEvo, giving us a concrete trick to steal for managing our own RL buffers during evolutionary search.

#2Efficient LLM Serving for Agentic Workflows: A Data Systems Perspective

MUST-READ

2026-03-17 | National University of Singapore | 2603.16104

M=8 P=9 I=8

Helium optimizes LLM serving for batch agentic workflows by modeling them as query plans and using a Templated Radix Tree (TRT) to enable proactive KV caching and cache-aware scheduling. The results are rigorously backed by numbers, demonstrating up to 1.56x speedups over state-of-the-art systems (vLLM, Parrot, KVFlow) on complex multi-agent workflows, and the authors even validate their greedy scheduler's optimality gap against an MILP solver. The most valuable takeaway is the TRT abstraction, which captures global prefix hierarchies across a DAG of LLM calls to maximize KV cache reuse, rather than relying on reactive, per-call caching. This is highly actionable for us: we should directly examine their scheduling formulation and TRT implementation to improve resource allocation and memory management in our GPUSched and HERMES projects.

#3IEMAS: An Incentive-Efficiency Routing Framework for Open Agentic Web Ecosystems

MUST-READ

2026-03-18 | Shanghai Jiao Tong University | 2603.17302

M=7 P=8 I=7

IEMAS routes client requests to distributed LLM agents by formulating the assignment as a Min-Cost Max-Flow bipartite matching problem, using VCG auctions to align economic incentives with KV-cache reuse. The results are backed by solid vLLM simulations, demonstrating an 80.2% KV-cache hit rate and a 35% cost reduction over baselines like GraphRouter. The most actionable takeaway for us is their method of quantifying KV-cache affinity (via Longest Common Prefix) and embedding it directly as a cost-reduction weight in a network flow optimization model. While the decentralized VCG auction mechanics might be overkill for centralized clusters, we should absolutely steal their cache-aware MCMF formulation for our OR-based LLM inference scheduling (GPUSched) work.

#4inference-fleet-sim: A Queueing-Theory-Grounded Fleet Capacity Planner for LLM Inference

MUST-READ

2026-03-17 | MBZUAI, McGill University, University of Chicago, Tensormesh Inc | 2603.16054

M=7 P=8 I=8

This paper introduces a two-phase capacity planner (M/G/c analytical sweep followed by discrete-event simulation) to find minimum-cost GPU fleet configurations for LLM inference under strict latency SLOs. The results are rigorously backed by simulation across multiple GPU profiles and workload traces, demonstrating that intuitive sizing rules often fail (e.g., slower GPUs can be cheaper due to KV-slot multipliers, and analytical models approve broken fleets for high-variance agent traffic). The most actionable takeaway for us is their lightweight, physics-informed (W, H, nmax) linear roofline model for GPU performance, which we can directly extract and embed into our integer programming formulations for GPUSched. This is highly relevant for our OR-for-AI infrastructure work, and we should use their open-source workload CDFs and performance constants to benchmark our own scheduling algorithms.

#5Guaranteeing Semantic and Performance Determinism in Flexible GPU Sharing

DISCUSS

2026-03-17 | Shanghai Jiao Tong University, Chinese Academy of Sciences, University of Chinese Academy of Sciences | 2603.15042

M=4 P=8 I=6

DETSHARE introduces 'GPU coroutines' to decouple logical execution contexts from physical GPU resources, enabling fine-grained spatial sharing without modifying kernels to preserve semantic determinism. The results are highly credible and backed by strong empirical numbers on A800/Hopper GPUs, demonstrating up to 79% higher training throughput and 69% lower inference latency compared to temporal sharing baselines. The most useful takeaway for us is the identification of 'semantic determinism'—specifically how dynamic spatial partitioning alters floating-point reduction trees and ruins training/RLHF stability. While we approach GPU scheduling via OR formulations (MIP/queueing) rather than OS-level CUDA hacking, this paper matters for our GPUSched project. We must incorporate their identified system realities—such as the 12% preemption overhead and the strict requirement for semantic determinism—into our mathematical models to ensure our theoretical schedules are actually deployable in production.

#6Cost-Efficient Multimodal LLM Inference via Cross-Tier GPU Heterogeneity

MUST-READ

2026-03-13 | University of Illinois Urbana-Champaign | 2603.12707

M=5 P=8 I=8

Yu et al. demonstrate that partitioning multimodal LLM inference at the modality boundary (vision encoder vs. language decoder) reduces cross-device transfer costs by O(L), dropping requirements from GB-scale NVLink to MB-scale PCIe. This enables heterogeneous serving architectures where cheap compute-dense GPUs (RTX 4090) handle vision and expensive bandwidth-dense GPUs (A100) handle language. Results are strongly backed by real hardware deployments, showing a 37% improvement in cost-efficiency over homogeneous vLLM baselines. WHAT WE LEARNED: The phase-separable nature of MLLMs and the use of cross-tier work stealing (idle 4090s assisting with language decoding) are massive structural opportunities. We must immediately update our integer programming formulations in the GPUSched project to model this modality-level disaggregation, otherwise our OR models will be obsolete for multimodal workloads.