Finance is stress-testing LLMs and finding confident hallucinations
Four new preprints put LLMs through finance's gauntlet: 15-23% of confident answers are wrong on FinQA, a 1,575-instance benchmark grades models by backtest rather than syntax, and proprietary models still lead open-weight ones 92.8% to 84.1% on accounting-relation extraction.
- In a preprint, the authors report that 15-23% of confident answers (where 8 resamples agree) are wrong on the FinQA benchmark, and that linear probes on the residual stream significantly outperform output-level baselines (log-probabilities and self-assessment) in detecting confident hallucinations. Self-reported m:…3827
- In a preprint, QuantEval's authors describe a 1,575-instance expert-validated benchmark across knowledge-based QA, quantitative mathematical reasoning, and quantitative strategy coding, graded with deterministic CTA-style backtesting rather than syntax or template matching; they report domain-aligned SFT and RL improve quantitative reasoning but trade off factual recall. Self-reported m:…4784
- In a preprint, the authors report proprietary models outperform open-weight models in financial relation classification, achieving 92.8% accuracy compared to 84.1%, on schema-guided evidence discovery over SEC 10-K filings. Self-reported m:…4366
- In a preprint, AlphaRLM's authors report it outperforms strong factor-mining and LLM-based baselines in annualized excess return across CSI 300, CSI 1000, and S&P 500. Self-reported m:…9128
Confidently wrong
The most alarming number in this batch: 15-23% of confident answers — defined as cases where 8 resamples agree — are wrong on the FinQA benchmark. These are not hedged guesses; they are the answers the model is surest about. The paper's proposed defense: linear probes on the residual stream significantly outperform output-level baselines (log-probabilities and self-assessment) in detecting confident hallucinations. Read the model's insides, not its confidence. m:…3827
The 15-23% range is worth sitting with — our read. It means agreement across resamples — the standard cheap uncertainty filter — is screening out nothing. Consensus is not correctness, and in finance the confident wrong answer is the expensive one. m:…3827
Grade the strategy, not the syntax
A second preprint, QuantEval, argues that the way we grade quant models is broken: syntax or template matching cannot tell you whether a generated strategy works. Its replacement is a 1,575-instance expert-validated benchmark across knowledge-based QA, quantitative mathematical reasoning, and quantitative strategy coding — and it grades the coding tasks with deterministic CTA-style backtesting, which the authors call a more effective evaluation method for financial strategy coding than syntax or template matching. The training finding: domain-aligned SFT and RL improve quantitative reasoning in LLMs but result in trade-offs regarding factual recall. m:…4784
The factual-recall trade-off is the paper's sting — our read. The same domain training that makes the model better at quantitative reasoning makes it worse at remembering facts. Specialization has a price, and in finance the price is paid in the exact currency the first paper is warning about: confident wrongness. m:…4784 m:…3827
The accounting layer
A third preprint moves from generic QA to the structure of the filings themselves. Financial disclosures contain underlying double-entry accounting relationships that can be extracted using schema-guided evidence discovery, the authors write — and their benchmark tests exactly that. The headline gap: proprietary models outperform open-weight models in financial relation classification, achieving 92.8% accuracy compared to 84.1%. The authors also report substantial variation in extraction volume and edge distributions across different LLM families when applied to SEC 10-K filings — the models do not even agree on how much structure they see. m:…4366
Factors from the literature
The fourth preprint, AlphaRLM, automates symbolic factor discovery by aligning LLM generation with financial literature: Literature-Guided Operator Assembly (LEGO) provides effective semantic priors for symbolic factor construction, and reinforcement learning with executable-validity and grouped-return-separation rewards aligns LLM generation policies with market feedback. The reported result: AlphaRLM outperforms strong factor-mining and LLM-based baselines in annualized excess return across CSI 300, CSI 1000, and S&P 500. This is research about evaluation and generation methods — not a strategy to trade. m:…9128
Our read
Our read: finance is building its own LLM eval stack, and it looks nothing like the generic leaderboards — probes for hallucinations, deterministic backtests for generated strategies, schema-guided extraction for filings. The pattern: in a domain where a wrong number has a dollar sign attached, the field is converging on evaluation that mirrors the actual work. m:…3827 m:…4784 m:…4366
The 92.8-vs-84.1 gap is also a reminder — our read — that in finance, unlike in chat, the open-weight frontier still trails, and the gap is measured on the boring, structured work that actually matters. The honest caveat: all four are preprints; backtests are not live results, and the 15-23% figure is on a benchmark, not a trading desk. m:…4366 m:…4784 m:…3827
What is still unresolved
- Whether linear-probe hallucination detection holds up in production financial QA pipelines, not just on FinQA.
- How QuantEval's deterministic CTA-style backtesting maps to live trading conditions — backtests are not live results.
- All four are preprints with no independent replication yet.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.