PHITSBench: an execution-scored benchmark for AI-assisted PHITS radiation-transport input generation using natural language

Autonomous language-model agents are increasingly evaluated on long-horizon tool-use tasks, but existing benchmarks rarely capture the complexity and nuance of real scientific work. To address this gap, we introduce Collider-Bench, a benchmark for evaluating whether LLM agents can reproduce experimental analyses from the Large Hadron Collider (LHC) using only public papers and open scientific software. Such analyses are often difficult to reproduce because the public toolchain only approximates the software used internally by the experimental collaborations, while the published papers inevitably omit implementation details needed for a faithful reconstruction. Agents must therefore rely on physical reasoning, domain knowledge, and trial-and-error to fill these gaps. Each task requires the agent to turn a published analysis into an executable simulation-and-selection pipeline and submit predicted collision event yields in specified signal regions. These predictions are evaluated with standard histogram metrics that provide continuous fidelity scores without a hand-written rubric. We also report the computational cost incurred by each agent per task. Finally, we evaluate the codebase and full session trace using an LLM judge to catch qualitative failure modes such as fabrications, hallucinations and duplications. We release an initial set of tasks drawn from LHC searches, together with a containerized sandbox and event simulation tools. We evaluate across a capability ladder of general purpose coding agents. Our results show that on average no agent reliably beats the physicist-in-the-loop solution.

Small language models (1-3B) are practical to run locally, but individually limited on harder code generation tasks. We ask whether composing them into pipelines can recover some of that lost capability. We study code generation pipelines built from 1-3B models with execution feedback, and use a NEAT-inspired evolutionary search to test whether more complex pipeline structure helps beyond a simple refinement loop. We evaluate on HumanEval (164 problems) and sanitized MBPP (427 problems), all with local inference on a single laptop. Self-refinement with execution feedback improves code generation by more than 4 standard deviations on both benchmarks. The gains are narrow in mechanism: refinement fixes many runtime errors (especially NameError and SyntaxError), but rarely fixes logic errors such as AssertionError. Within our tested general-purpose model pool, generator identity mattered less than refiner capability: a 1.5B generator paired with a 3B refiner matched a 3B model doing both roles. Early stopping is essential; without it, every iteration is net-negative. The code-specialized models outperform every general-purpose pipeline configuration, suggesting model specialization matters more than pipeline architecture. Preliminary text-only pipeline experiments without execution feedback did not show gains at this scale. In our constrained search space, evolutionary search mostly rediscovered the same simple generate-execute-refine loop we found manually, with no clearly significant gain from added topology. Single-evaluation fitness inflates results by 5-7 percent, selecting lucky genomes over good ones. On these benchmarks at 1-3B scale, execution feedback mattered more than added pipeline complexity in determining whether composition helped.

When a hint turns a failing generated program into a passing one, does it provide missing information or merely steer the model toward a solution it could already produce? We test these hypotheses on HumanEval+ and MBPP+ using executable evaluation. For Qwen2.5-3B-Instruct, adaptive relevant hints rescue 36 of 79 selected failures; an unrelated hint rescues 19, while eight unhinted samples solve 46 and recover 31 of the 36 relevant-hint rescues. Phi-3.5-mini shows the same pattern: relevant hints rescue 42 of 101 failures, an unrelated hint rescues 17, and unhinted sampling solves 57, including 36 of the 42 relevant-hint rescues. Because the hint conditions use different attempt budgets, these comparisons do not isolate a purely semantic effect. Mechanistic tests on Qwen identify a stable activation direction shared by relevant and unrelated hints. Persistently adding this direction yields 14 rescues and 18 regressions, with no detectable net accuracy gain; learned low-rank interventions have a positive but imprecise estimated effect. Full textual specifications solve 22 of 24 context-defined problems, versus 5-11 for tested virtual-KV prefixes. Post-generation hidden-state probes transfer across benchmarks, with pooled AUROC 0.806 and 0.780, but their top-one selection advantage over token confidence is statistically unresolved. Overall, relevant hints can rescue failures, but most rescued solutions are already reachable through ordinary sampling, and the internal interventions tested here do not establish task-general capability transfer.