RI

#1OptiRepair: Closed-Loop Diagnosis and Repair of Supply Chain Optimization Models with LLM Agents

MUST-READ

2026-02-23 | Massachusetts Institute of Technology, Alibaba Group | 2602.19439

M=8 P=7 I=9

Ao et al. introduce OptiRepair, a closed-loop framework that repairs infeasible LPs using solver IIS feedback (Phase 1) and validates them with a 'Rationality Oracle' based on domain theory (Phase 2). Results are exceptionally strong: fine-tuned 8B models trained via iterative STaR and GRPO achieve 81.7% success, outperforming GPT-5.2 (42.2%) by a massive margin. Key Takeaway: We should steal the 'Rationality Oracle' concept—evaluating solution properties (e.g., monotonicity, variance bounds) rather than just raw fitness—to serve as a dense signal for our Process Reward Models in AlgoEvo. Additionally, their success with solver-verified GRPO confirms we should prioritize training specialized operators over prompting general LLMs.

#1AdaEvolve: Adaptive LLM Driven Zeroth-Order Optimization

MUST-READ

2026-02-23 | University of California, Berkeley, Bespoke Labs | 2602.20133

M=9 P=9 I=9

AdaEvolve replaces static evolutionary schedules with a three-tier adaptive controller: local exploration intensity based on an 'accumulated improvement signal' (pseudo-gradient), global compute allocation via a normalized bandit, and meta-level 'tactic' generation when stagnation occurs. Results are highly convincing, showing SOTA on Circle Packing (beating AlphaEvolve) and 185 other tasks while using the same LLM backbone as baselines, proving the gains are algorithmic. The most stealable insight is the $G_t$ signal metric—an exponential moving average of squared normalized improvements—which allows auto-tuning exploration rates without manual intervention. This is a direct upgrade for our AlgoEvo architecture, specifically addressing our sample efficiency and stagnation bottlenecks.

#1DualPath: Breaking the Storage Bandwidth Bottleneck in Agentic LLM Inference

MUST-READ

2026-02-25 | DeepSeek-AI, Peking University, Tsinghua University | 2602.21548

M=6 P=9 I=7

Wu et al. introduce DualPath, a system that breaks the storage I/O bottleneck in agentic inference by utilizing idle decode-node bandwidth to load KV-cache and transferring it to prefill nodes via RDMA. The results are highly credible (1.87x throughput increase on DeepSeek-V3) and directly target the long-context, short-append patterns typical of our evolutionary search rollouts. The most valuable technical takeaway is the 'CNIC-centric' traffic management strategy, which isolates bulk KV transfer from latency-sensitive model collectives—a technique we should immediately steal for our own serving infrastructure. While their scheduling logic is a simple heuristic, the architectural change defines a new, more flexible state space for our OR-based resource allocation research.

#2Training Generalizable Collaborative Agents via Strategic Risk Aversion

MUST-READ

2026-02-25 | Caltech | 2602.21515

M=8 P=9 I=8

Qu et al. introduce Strategically Risk-Averse Policy Optimization (SRPO), which trains agents against a 'constrained adversary' that minimizes their reward within a KL-divergence bound of the partner's current policy. Theoretical results prove this objective eliminates free-riding equilibria, and experiments on GSM8K multi-agent debate show it prevents 'lazy' agreement, improving joint accuracy by up to 19% when pairing heterogeneous LLMs (e.g., 0.6B with 4B). The key takeaway is that robustness to partner deviation—enforced via this specific adversarial objective—is a more principled way to fix lazy agent behavior than prompt engineering or simple dropout. We should immediately test this objective in our HERMES debate framework to improve the contribution quality of smaller models.

#3How to Allocate, How to Learn? Dynamic Rollout Allocation and Advantage Modulation for Policy Optimization

MUST-READ

2026-02-22 | Meituan, Tsinghua University, Fudan University, Peking University | 2602.19208

M=8 P=7 I=9

DynaMO introduces a dual-pronged optimization for RLVR: a dynamic rollout allocation strategy that prioritizes problems with high gradient variance (proxied by Bernoulli variance of success/failure), and a gradient modulation technique to stabilize updates. The results are strong (+11.8% Pass@1 over GRPO on Qwen-7B) and backed by clear ablations. The critical takeaway for us is the allocation logic: we should immediately replace uniform sampling in AlgoEvo with variance-based allocation ($n_i \propto \sqrt{p(1-p)}$). This ensures compute is spent on instances/components that are currently 'learnable' (high variance) rather than wasted on the trivial or impossible, directly optimizing our search budget.

#4ARLArena: A Unified Framework for Stable Agentic Reinforcement Learning

MUST-READ

2026-02-25 | University of California, Los Angeles, University of Wisconsin–Madison | 2602.21534

M=8 P=8 I=8

The authors dissect why standard RL (GRPO/PPO) fails in multi-turn agentic tasks, identifying that token-level importance sampling (IS) clipping allows negative-advantage outliers to destabilize training. They propose SAMPO, which enforces sequence-level clipping and integrates fine-grained step-level advantages (similar to process rewards) to stabilize learning. The results are rigorous, showing a jump from ~50% to 92% success on ALFWorld by fixing the gradient update mechanics rather than just prompt engineering. Key Takeaway: We must audit our RL implementations; if we are using token-level clipping for multi-step evolutionary agents, we are likely suffering from silent gradient instability—switching to sequence-level clipping and masking negative-advantage outliers is an immediate, code-level improvement we should adopt.

#5Reasoning-Driven Design of Single Atom Catalysts via a Multi-Agent Large Language Model Framework

DISCUSS

2026-02-25 | Georgia Institute of Technology, Korea University, Sogang University, Ewha Womans University | 2602.21533

M=7 P=3 I=8

Mok et al. propose MAESTRO, a multi-agent LLM framework for optimizing single-atom catalysts that explicitly separates search into exploration and exploitation phases, bridged by a textual 'Exploration Report.' Results are validated against high-fidelity DFT calculations, showing the system learns to break theoretical scaling relations via in-context learning, outperforming memory-less baselines. The key takeaway for us is the Exploration Report Agent: instead of just passing the best candidates to the exploitation phase, the system pauses to write a natural language strategy guide summarizing 'what worked and what didn't' from the exploration phase. We should steal this mechanism for AlgoEvo to let the search agent 'learn' from the initial population generation rather than just selecting from it.

#6GauS: Differentiable Scheduling Optimization via Gaussian Reparameterization

DISCUSS

2026-02-23 | Cornell University, University of Maryland, College Park | 2602.20427

M=7 P=5 I=7

GauS replaces the standard categorical (Gumbel-Softmax) relaxation in differentiable scheduling with Gaussian variables defined by mean and variance, reducing parameter space from O(N*D) to O(N). Results are strong: it scales to 57k nodes where previous differentiable methods OOM and exact solvers timeout, while maintaining near-100% GPU utilization. The key takeaway is a specific modeling technique: using Gaussian distributions to represent discrete ordinal values (like time steps) naturally captures temporal proximity and provides smoother gradients than categorical buckets. We should test this representation in our continuous latent-space optimization work to replace categorical relaxations for ordered parameters.