LamBench turns lambda calculus into coding test

Sharp idea, but also a reminder that new benchmarks can be most useful before models get tuned to them. This one rewards symbolic reasoning, syntactic discipline, and exact execution more than familiar code-generation muscle.

• –The task set is unusually deep for a niche benchmark: Church and Scott encodings, lists, trees, ADTs, plus harder algorithms like SAT, FFT, Sudoku, and TSP.
• –Scoring is straightforward pass rate, with solution size tracked as a secondary metric, which makes the results easy to read and harder to hand-wave.
• –The current leaderboard is the real story: GPT-5.4 leads GPT-5.5, and the gap to mid-tier and open models is large, which suggests the benchmark is stress-testing a very specific skill mix.
• –Because the benchmark is still new and intentionally simple, the scores should be treated as a directional signal, not a broad verdict on general coding ability.
• –For teams working on reasoning-heavy agents or compiler-like synthesis, this is a useful addition to the eval stack; for ordinary app coding, it is probably too synthetic to stand alone.