A benchmark score says how well an AI did, but can saved logs prove it actually played fair?
Can infrastructure evidence ground benchmark claims better than terminal scores alone?
This explores whether recording how an agent actually reached its result, through logs and infrastructure records, gives a more trustworthy basis for benchmark claims than the final score alone.
This explores whether recording how an agent reached its result gives a firmer basis for benchmark claims than the final score alone. The corpus says yes, and it comes from several directions. The most direct case is BenchShield. It lets a benchmark operator issue a claim about valid completion, meaning the agent followed the intended evaluation path, and back that claim with recorded infrastructure evidence. Can infrastructure evidence replace terminal scores in benchmark validation? The output changes from a number to a verifiable statement. A score says how well the agent did, while a claim says whether the result counts.
Other notes show why a score alone can mislead, without proposing infrastructure evidence as the fix. Qwen2.5-Math-7B can reconstruct 54.6% of MATH-500 from partial prompts yet scores 0.0% on a post-release benchmark. Its high score reflected memorization, not reasoning. Does RLVR success on math benchmarks reflect genuine reasoning improvement? Real reasoning behavior and inflated benchmark gains can sit side by side in the same model, so the number can't tell you which one you're looking at. Can genuine reasoning activation coexist with contaminated benchmarks? Cybersecurity benchmarks have a similar gap. Models score well on vulnerability reproduction, patching and CTF tasks, but exploitation, the step where a flaw becomes a real attack, is barely measured. Do cybersecurity benchmarks actually measure exploitation? A single score also hides differences across capabilities. Models that rank first on task success often rank lower on privacy compliance or long-horizon retention. Does a single benchmark score actually predict agent readiness?
There is an apparent conflict. One note argues reasoning benchmarks should score final answers, not reasoning traces, because trace scoring rewards stylistic mimicry of reasoning. Should reasoning benchmarks score final answers or reasoning traces? A related result found that logically invalid chain-of-thought examples work almost as well as valid ones, so reasoning-shaped text is a poor sign of real inference. Does logical validity actually drive chain-of-thought gains? The two positions can be reconciled if you separate the kinds of evidence. Judging how convincing the reasoning text looks is unreliable. Recording what happened in the environment is a different thing. On my reading of the summaries, infrastructure evidence is the second kind: it checks the path, not the prose.
Evidence collection also helps where judging is involved. An agentic evaluator that gathers evidence dynamically had a 0.27% judge shift, against 31% for a plain LLM judge on complex tasks. But its memory module cascaded errors, so it needs error isolation to keep the gain. Can agents evaluate AI outputs more reliably than language models? Evidence also doesn't settle the comparison problem. Interactive and trajectory-level evaluations move the old problems of comparability, reproducibility and mapping evidence to a judgment into a higher-dimensional space. They don't remove them. Do interactive evaluations actually solve the benchmark comparison problem? Grounding claims in infrastructure evidence is therefore stronger than trusting a terminal score. It only becomes comparable across labs if the field also agrees on shared standards for what counts as valid evidence.
Sources 9 notes
BenchShield enables benchmark operators to issue claims about valid task completion grounded in recorded infrastructure evidence rather than terminal scores alone. This shifts from a single number to a verifiable claim about whether an agent followed the intended evaluation path.
Qwen2.5-Math-7B reconstructs 54.6% of MATH-500 from partial prompts but scores 0.0% on post-release LiveMathBench, revealing dataset contamination. On clean benchmarks, only correct rewards improve performance; random and inverse rewards fail or degrade reasoning ability.
RLVR activates genuine reasoning patterns through RL training while benchmark improvements may reflect data memorization on contaminated datasets. These operate at different measurement levels and can coexist without contradiction.
ExploitGym shows that while frontier models excel at vulnerability reproduction, patch generation, and CTF tasks, exploitation—the step where a vulnerability becomes a real attack—remains largely unmeasured in the benchmark literature.
Capability decomposes into task success, privacy compliance, long-horizon retention, mode-shift behavior, and ecosystem readiness. Models ranked highest on one axis often rank lower on others, making single-score evaluations systematically misleading for real deployment.
Show all 9 sources
LR²Bench scores only final answers against deterministic ground truth, not reasoning steps. This methodological choice reveals a 20% ceiling that trace-based evaluation would inflate by counting stylistic reasoning mimicry as actual reasoning capability.
Illogical chain-of-thought exemplars matched valid CoT performance on BIG-Bench Hard, showing that structural properties—not logical validity—drive the gains. The model learns the form of reasoning, not genuine inference.
Eight-module agentic evaluation achieved 0.27% judge shift versus 31% for LLM-as-a-Judge on complex tasks. However, the memory module cascaded errors, revealing that agentic systems need error isolation mechanisms to maintain gains.
Interactive evaluation relocates core problems—comparability, reproducibility, evidence-to-judgment mapping—into higher-dimensional space rather than solving them. The field needs design protocols and shared standards, not format adoption, to make trajectory scoring interpretable.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Local Coherence or Global Validity? Investigating RLVR Traces in Math Domains
- Spurious Rewards: Rethinking Training Signals in RLVR
- Interactive Evaluation Requires a Design Science
- The Invisible Leash: Why RLVR May Not Escape Its Origin
- Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?
- Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate
- Hack-Verifiable Terminal Bench: Evaluating Reward Hacking in Terminal Tasks
- Cyber-Capable AI Agents: Vulnerabilities, Evaluation Containment, and Defensive Response