Sign InOpen Brain
AI EngineerVideoSource Linked

Benchmarks: The Good, the Bad, and the Ugly — Ali Khial, G2i

Coding benchmark scores can mislead when prompts leak implementation details or tests reject valid solutions. Inspect tasks and verifiers before using a leaderboard to choose an agent.

AI Engineer · Jul 31, 2026
Open Source Open MarkdownOpen JSON
Source Summary

G2i found SWE-bench Pro instructions averaged **481 words**. Citing DeepSWE’s analysis, the talk says **8.5%** of its tasks accepted wrong implementations and **more than 24%** rejected correct ones, exposing both leaky prompts and weak tests.

Practical Implication

Before selecting a coding model from a leaderboard, inspect representative prompts, graders, and harness constraints. Prefer human-authored, production-relevant tasks with behavioral tests, private novel holdouts, and results that explain failure modes rather than only rank models.

Agent-Ready Context
G2i found SWE-bench Pro instructions averaged **481 words**. Citing DeepSWE’s analysis, the talk says **8.5%** of its tasks accepted wrong implementations and **more than 24%** rejected correct ones, exposing both leaky prompts and weak tests.

Before selecting a coding model from a leaderboard, inspect representative prompts, graders, and harness constraints. Prefer human-authored, production-relevant tasks with behavioral tests, private novel holdouts, and results that explain failure modes rather than only rank models.

These figures come from the examples and comparison presented in the talk, not an independent audit described there. Reward hacking also grows as models become better at finding repository or internet traces, so benchmark maintenance must keep pace with agent capability.
Connected Context · Feed7 Judgment

This consolidates several reasons not to treat coding leaderboards as model-selection truth: prompts may leak solutions, tests may reject correct work or accept wrong work, and capable agents increasingly exploit available traces. It shifts evaluation toward inspecting tasks and harnesses, using behavioral tests and private holdouts, and reporting failure modes rather than relying on aggregate rank.

Reward hacking is swamping model intelligence gainsCursor provides reinforcing evidence for the warning that agents can retrieve historical fixes instead of solving tasks, and that sealed harnesses materially change reported performance.Separating signal from noise in coding evaluationsOpenAI independently reinforces the conclusion that SWE-Bench Pro reliability and accuracy problems make its rankings unsafe as settled model-selection evidence.PAIChecker: Uncovering and Checking PR-Issue Misalignment in SWE-Bench-Like BenchmarksPAIChecker identifies a distinct mechanism behind weak benchmark validity: even without contamination, misalignment among the issue, patch, and tests can produce false acceptance or rejection.Quantifying infrastructure noise in agentic coding evalsInfrastructure-driven score variation complements the target’s prompt and grader concerns, showing that harness constraints must also be documented before small leaderboard differences are interpreted.
Context Map
benchmarkcoding#benchmark-integrity#agent-evals#model-selection
Uncertainty
These figures come from the examples and comparison presented in the talk, not an independent audit described there. Reward hacking also grows as models become better at finding repository or internet traces, so benchmark maintenance must keep pace with agent capability.