Skip to content
study-swarm

Research grounding

study-swarm is itself research-grounded. Every load-bearing choice in the verification gate traces to published evidence. (Naturally, these citations were themselves verified by the protocol — see the proof at the bottom.)

LLMs can’t reliably self-verify

Section titled “LLMs can’t reliably self-verify”
  • Huang et al. 2023 (arXiv:2310.01798) — LLMs struggle to self-correct without external feedback; performance can even degrade.
  • Kambhampati et al. 2024 (arXiv:2402.01817, LLM-Modulo) — pair the generator with an external model-based verifier; autoregressive LLMs can’t self-verify.
  • Stechly et al. 2024 (arXiv:2402.08115) — the external verifier carries the gains; the self-critique content is largely inert.

Same-family judges self-prefer

Section titled “Same-family judges self-prefer”
  • Panickssery, Bowman & Feng 2024 (arXiv:2404.13076) — self-recognition correlates linearly with self-preference, so partial blinding doesn’t help.
  • Verga et al. 2024 (arXiv:2404.18796, PoLL) — a panel across disjoint families has less bias at ~7× lower cost than one large judge.

Citations are where LLMs lie — so check by retrieval

Section titled “Citations are where LLMs lie — so check by retrieval”
  • Walters & Wilder 2023 (doi:10.1038/s41598-023-41032-5) — 55% of GPT-3.5 / 18% of GPT-4 citations are fabricated; many real ones carry bibliographic errors.
  • Onweller et al. 2026 (arXiv:2605.06635) — cited links resolve >94% of the time, yet only 39–77% of the content actually supports the claim. “The link resolves” ≠ “the paper says this.”

Hide the generator’s reasoning

Section titled “Hide the generator’s reasoning”
  • Khalifa et al. 2026 (arXiv:2601.14691, “Gaming the Judge”) — manipulated chain-of-thought alone can inflate a judge’s false-positives by up to 90% with actions held fixed.
  • Turpin et al. 2023 (arXiv:2305.04388) — CoT is often a post-hoc rationalization, not the real decision path.

Diversity beats count

Section titled “Diversity beats count”
  • Rajan 2025 (arXiv:2511.16708) — four verifiers at pairwise correlation ρ ∈ [0.05, 0.25] beat any single one via submodular coverage. (Anchored in code verification; the submodular argument is domain-general.)
  • Kim et al. 2025 (arXiv:2506.07962) — LLM errors are correlated, so lens diversity, not raw count, is the load-bearing variable.

Groundedness is a distinct, measurable axis

Section titled “Groundedness is a distinct, measurable axis”
  • Min et al. 2023 (arXiv:2305.14251, FActScore) and Gao et al. 2023 (arXiv:2305.14627, ALCE) — even strong models fail to fully support their own citations roughly half the time.

The proof

Section titled “The proof”

As a test, the protocol was run against its own citations. Two decorrelated non-Claude families — Mistral (mistral-small:24b) and IBM Granite (granite4.1:30b) — checked a citation set, reasoning-stripped, seeded with two blind traps:

Planted trapMistralIBM GraniteTruth
CoT prompting attributed to “Nakamura & Olsen”missedcaughtmisattributed (really Wei et al. 2022, arXiv:2201.11903)
a fabricated “98% of errors removed” papercaughtcaughtfabricated

Neither family caught both alone — their union caught 2/2. A single judge would have shipped the misattribution. Meanwhile the retrieval oracle caught two real misattributions in adjacent design docs and correctly confirmed genuine 2026 papers that both LLMs false-flagged as fabricated (because the papers postdate their training) — which is exactly why the existence check must be a retrieval oracle, never an LLM.