RI

#1Rethinking Code Similarity for Automated Algorithm Design with LLMs

2026-03-03 | City University of Hong Kong |2603.02787

Zhang et al. propose BehaveSim, a metric that measures algorithmic similarity by applying Dynamic Time Warping (DTW) to the sequence of intermediate solutions (trajectories) generated during execution, rather than relying on static code analysis. By integrating this into FunSearch and EoH to enforce behavioral diversity, they achieve significant performance gains, notably reducing the optimality gap on TSP by ~7.8% compared to standard FunSearch.Key Takeaway:

#2Learning to Evolve for Optimization via Stability-Inducing Neural Unrolling

2026-03-03 | The Hong Kong Polytechnic University, The University of Hong Kong |2512.11453

Gao et al. propose L2E, a meta-learned neural optimizer that uses Mamba blocks to parameterize evolutionary operators within a stability-enforcing unrolled loop (Krasnosel'skii-Mann iteration). Results on BBOB and LSGO-1000D are strong, showing it outperforms Transformer-based L2O methods (GLHF) and classical heuristics (DE) in sample efficiency and zero-shot generalization.Key Takeaway:

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

2026-03-05 | Kingsoft Cloud |2603.04716

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.{eff} = TP{max} - \frac{L{in}}{TTFT - T{overhead}}$).

#2AI-for-Science Low-code Platform with Bayesian Adversarial Multi-Agent Framework

2026-03-03 | Fudan University, Shanghai Innovation Institute, Shanghai Academy of AI for Science |2603.03233

The authors propose a multi-agent framework for scientific code generation that couples an adversarial 'Challenger' (generating difficult test cases) with a 'Solver', governed by a Bayesian update rule. Crucially, they employ Bayesian Optimization with a kernel based on code embeddings (AST + text) to estimate solution qualitybeforerunning expensive tests, effectively acting as a learned surrogate model. Results on SciCode and ScienceAgentBench are strong, showing small models (Qwen-32B) outperforming GPT-4o when using this loop.

#3VisionCreator: A Native Visual-Generation Agentic Model with Understanding, Thinking, Planning and Creation

2026-03-03 | Tencent Hunyuan, Hong Kong University of Science and Technology |2603.02681

This paper introduces VisionCreator, an agent trained via 'Virtual Reinforcement Learning' (VRL) where tool outputs and logic are simulated to train long-horizon planning policies without incurring expensive real-world execution costs. They employ a 'Plan-Driven Reward' model (combining LLM-based plan verification with rule-based execution checks) and prove theoretical bounds for the sim-to-real transfer, achieving performance superior to GPT-5 on visual tasks.Key Takeaway:

#4StitchCUDA: An Automated Multi-Agents End-to-End GPU Programing Framework with Rubric-based Agentic Reinforcement Learning

2026-03-03 | University of Minnesota-Twin Cities |2603.02637

StitchCUDA automates end-to-end GPU program generation using a multi-agent framework, but its core contribution is a training recipe that solves reward hacking in code optimization. They decompose expensive multi-turn agentic RL into single-turn 'atomic skills' (generation vs. refinement) and use GRPO with an LLM-evaluated 'Rubric Reward' (e.g., 'Did you use tiling?') rather than just sparse outcome metrics. This prevents the model from gaming the system (e.g., wrapping PyTorch code) and forces actual optimization behavior.

#5PromptTuner: SLO-Aware Elastic System for LLM Prompt Tuning

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

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.

#6BandPO: Bridging Trust Regions and Ratio Clipping via Probability-Aware Bounds for LLM Reinforcement Learning

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

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.

#7AI4S-SDS: A Neuro-Symbolic Solvent Design System via Sparse MCTS and Differentiable Physics Alignment

2026-03-04 | Nanjing University, Suzhou Laboratory, Shanghai Artificial Intelligence Laboratory |2603.03686

Chen et al. introduce a neuro-symbolic MCTS framework for mixed discrete-continuous optimization, applying it to solvent design. They solve the LLM context bottleneck via 'Sparse State Storage' (storing only state abstractions and reconstructing paths on-demand) and fix mode collapse using 'Sibling-Aware Expansion' (conditioning the generator on sibling nodes to force orthogonality).

#8Build, Judge, Optimize: A Blueprint for Continuous Improvement of Multi-Agent Consumer Assistants

2026-03-03 | DoorDash, WithMetis.ai |2603.03565

This paper presents a production-grade framework for optimizing multi-agent systems by jointly evolving prompt bundles (MAMUT) rather than optimizing agents in isolation. They validate this on a grocery assistant, showing that system-level optimization outperforms local sub-agent optimization by ~7% because it captures coordination dynamics (e.g., context passing) that local metrics miss. The most stealable insight is their 'Judge Calibration' loop: they use evolutionary search (GEPA) to optimize theevaluator'sprompt to match human labels (91.4% agreement) before using that judge to optimize the agents.

#9MuxTune: Efficient Multi-Task LLM Fine-Tuning in Multi-Tenant Datacenters via Spatial-Temporal Backbone Multiplexing

2026-03-03 | Shanghai Jiao Tong University, National University of Singapore |2603.02885

MuxTune introduces a hierarchical scheduler for multi-tenant PEFT that uses Dynamic Programming to optimally fuse tasks (spatial batching) or interleave them (temporal multiplexing) based on a pipeline cost model. Empirical results on H100s show up to 5x throughput gains over NeMo and S-LoRA, validated by ablation studies. The most stealable insight is theirchunk-based data alignment