SciNovelty-Bench (seed)

A contamination-aware, verifiable benchmark for measuring whether LLMs productively generalize to genuinely novel scientific and mathematical problems — rather than retrieving memorized patterns.

Author: Kedir Yassin Hussen · PhD candidate, AI & Robotics, Bahir Dar University

This repository is the public seed of SciNovelty-Bench. It contains a working, code-verified synthetic core plus worked examples of every control tier, the scoring code, and the construction protocol. It is an early-stage artifact shared to make the methodology inspectable; it is not yet the full benchmark.

Why this benchmark

Frontier models now score above 95% on MMLU, GSM8K, and HumanEval — these benchmarks have saturated and no longer separate top systems. Yet the same models still collapse on genuinely unfamiliar scientific reasoning. The core problem is that standard benchmarks measure retrieval of memorized content, and as the public problem space leaks into training data, high scores increasingly reflect memorization rather than reasoning.

SciNovelty-Bench attacks this with three commitments:

• Verifiable — every scored item has a deterministic, code-computed gold answer (no human-annotator bottleneck, no subjective grading).
• Contamination-aware — items are either synthetic (novel by construction) or sourced strictly from competition sittings that postdate each tested model's training cutoff. Every synthetic item also carries a recorded best-effort public-corpus check (corpus_check).
• Novelty vs. rarity — obfuscated controls separate genuine structural novelty from mere surface unfamiliarity, and nonsense controls separate meaningful novelty from gibberish.

Tier design

Tier	Purpose	Verification
familiar_control	Map the model's known-reasoning manifold (baseline)	numeric_exact / tolerance
verifiable_novel_synthetic	Problems that combine two or more known primitives in unprecedented ways; novel by construction	numeric_exact / tolerance
verifiable_novel_competition (planned)	Elite competition problems (IMO, Putnam, etc.) from sittings after each model's cutoff	numeric_exact / tolerance
obfuscated_control	Structure identical to a synthetic item, surface wording made unfamiliar/cross-domain — isolates structural generalization from surface rarity	numeric_exact
control_nonsense	Grammatically plausible but semantically void — high surface novelty, no valid answer	ill_posed (model should decline/flag)

The competition tier is intentionally empty in this seed: fabricating "post-cutoff" items would defeat the contamination guarantee. It will be sourced per model at evaluation time.

The 20 synthetic items span number theory, counting, modular arithmetic, combinatorics, recurrences, series, bitwise operations, base conversion, physics and chemistry.

Item schema

{
  "id": "syn-0001",
  "tier": "verifiable_novel_synthetic",
  "domain": "number_theory",
  "construction_method": "compositional",
  "primitives": ["digit_sum", "num_divisors"],
  "prompt": "Define f(n) = (sum of the decimal digits of n) x (number of positive divisors of n). Compute f(f(360)).",
  "gold_answer": 144,
  "verification": "numeric_exact",
  "tolerance": null,
  "novelty_basis": "novel-by-composition; parameters fixed at construction, not a known textbook/competition item",
  "contamination_status": "synthetic_generated_post_construction",
  "linked_item": null,
  "corpus_check": "web search 2026-10-15: no match found"
}

• linked_item connects an obfuscated control to the synthetic item whose computational structure it preserves.
• corpus_check records the public-corpus check for synthetic items (null for other tiers).

Public-corpus check

Before release, each synthetic prompt was searched on the public web (2026-10-15). Result titles were reviewed and likely matches were opened. Each item's corpus_check records the outcome: no match found, or the closest related problem and how it differs. Several items share a general template with known exercises (for example, sums of square divisors) but use different numbers or combinations; no exact instance was found. This is a best-effort check: the absence of a public match does not prove absence from any model's training data, and closed-model training data cannot be inspected.

Contents

• data/scinovelty_seed.jsonl — 29 items (20 verifiable synthetic, 4 familiar controls, 3 obfuscated controls, 2 nonsense controls).
• generate_seed.py — builds the seed; every gold answer is computed in code, not entered by hand.
• verify.py — the scorer (numeric_exact / numeric_tolerance / ill_posed) with a self-test confirming all 29 items pass their own verifier, plus structural checks (every synthetic item has at least two primitives and a corpus check; every obfuscated item matches its linked item's answer).

Reproduce / verify

python3 generate_seed.py     # regenerate data/scinovelty_seed.jsonl
python3 verify.py            # self-test: 29/29 items pass their own verifier

Research direction

This benchmark is Dataset 2 of my PhD research (Bahir Dar University), which reads a model's layer-wise internal states turn by turn, in English and Amharic. Two of its measures use these items:

• Novel Subspace Index (NSI): how far a turn's activations depart from a mapped region of known states (negative log-density under a per-layer density model; flagged above the 95th percentile of held-out known data).
• Alignment Order Parameter (AOP): whether activations lean toward aligned or harmful behaviour (cosine similarity to an aligned direction, with AOP-risk read against a harm direction).

The verifiable items here provide the external ground truth needed to separate genuinely novel reasoning from confident confabulation.

Status & roadmap

• Code-verified synthetic seed (20 items) + control tiers
• Scorer with self-test
• Best-effort public-corpus check for every synthetic item
• Scale synthetic suite (target: at least 80 items, at least 20 per tier); finalize per-model post-cutoff contamination verification
• Extend to five-turn trajectories and a parallel Amharic version
• Behavioral evaluation across frontier APIs and open models
• Layer-wise NSI/AOP activation analysis on open-weight families; holdout validation

Changelog

v0.2 (2026-10-15)

• Replaced three items that matched well-known textbook problems: syn-0010 (lattice paths), syn-0011 (falling stone) and syn-0014 (trailing zeros of 100!).
• Reworked nine single-primitive items so that every synthetic item combines at least two primitives: syn-0002, 0003, 0005, 0006, 0012, 0013, 0015, 0016 (now chemistry) and 0019. obf-0002 was rewritten to match the new syn-0013.
• Added the corpus_check field and the public-corpus check.
• verify.py now also runs structural checks.
• Research direction aligned with the PhD proposal: NSI is the Novel Subspace Index, and AOP is defined against aligned and harmful directions.

v0.1 — initial seed.

License

Code: MIT. Data: CC-BY-4.0.