Research Radar: Agents Beat Paid OR Consultants on Objective but Not on Deliverables, Two Papers Evolve the Harness With the Model, and Edge Kernels Come Out Correct but Slow

This edition covers four preprints submitted to arXiv between October 16 and 19, 2026 and listed on Hugging Face's Daily Papers on October 20 (the first three) and October 21 (EdgeKernelBench).

OR-License: real client work, real money (arXiv 2610.24087)

Most operations-research benchmarks start from a clean problem statement. OR-License starts from the client's own materials: 31 instances across 10 industrial scenarios drawn "from paid OR engagements totaling USD 3.68 million", with the original requirements, data and deliverable formats preserved after de-identification. A simulated client, ORSim, "discloses business facts without supplying OR expertise", and hidden checks compare each delivery with the expert solution the client originally accepted. The headline metric, Deployment-Adjusted Quality, scores infeasible deliveries as zero. Across 11 frontier models running in coding-agent harnesses, "several exceed the expert reference in conditional objective quality, but none matches its DAQ". The best Claude Code configuration reaches 88.39% strict feasibility and a DAQ of 0.87 against the experts' 1. Completion rates for the full chain of research steps range from 24% to 45%.

Self-HPO: let the policy rewrite its own harness (arXiv 2610.22791)

The authors argue that "a harness tuned for one checkpoint may not target the failure modes of an updated policy". Self-HPO alternates two steps. In the first, the current model analyses its own failures and proposes harness changes; in the second, it is trained with reinforcement learning under the chosen harness. The hard tasks from each round become the next harness search set. On three coding benchmarks it raises Qwen3.5-35B-A3B's average resolve rate from 57.8% to 64.8%, and the authors report that the joint gain "exceeds the sum of their individual gains". A stronger external harness proposer adds more.

HarnessPro: improve the harness, cut the cost (arXiv 2610.26770)

HarnessPro evolves harnesses without ground-truth labels. It scores candidates on LLM-generated tests and returns three of them, lite, medium and heavy, for different budgets. On Terminal-Bench 2, SWE-Bench Pro and ProgramBench with GPT-5.5 in Codex, the heavy mode averages 2.0 points more pass rate at 26.1% lower cost per task than the RHO baseline. The authors say the result also holds for GLM-5.2 in Claude Code.

EdgeKernelBench: correct is not fast (arXiv 2610.24648)

This benchmark targets edge hardware through a unified IR, using TVM as the reference compiler. With a single attempt, the best model reaches "64.5% Correct@1 but only 25.5% Fast@1.0". The authors' layered kernel agent, LEKA, reaches 43.5% Fast@1.0, against 13.0% for direct generation and 26.1% for Claude Code with the same base model and budget.

Analysis: OR-License is the most useful paper of the four for anyone deploying agents on client work. The gap between 'better objective when it works' and 'worse expected deliverable' comes from infeasible outputs, so feasibility checks against the client's real constraints matter more than solver quality. Self-HPO and HarnessPro treat the harness as a trainable part of the system, not fixed scaffolding. For teams that maintain their own harness, the practical step is to log failures in a form a model can read and propose fixes from. EdgeKernelBench repeats a lesson GPU kernel work already teaches: when generated code ships, require a performance gate alongside the correctness test.