Spotting an error in a given artifact

Claims

• mechanismsingle paper↑1↓0
  For best-of-N selection on mathematical reasoning, a small verifier (~2M parameters) that reads the frozen generator's own hidden states at reasoning-step boundaries ranks correct candidates above incorrect ones better than a larger text-only verifier that re-encodes the generated solution, because the correctness signal is already present in the generator's upper-layer representations and must otherwise be inferred indirectly from emitted tokens.
• observationsingle paper↑1↓0
  For detecting findings-conclusion inconsistencies in oncology PET/CT radiology reports, a 15M-parameter BERT pre-trained on in-domain reports reaches higher balanced accuracy and lower false-positive rate than zero- or few-shot prompted 27B-70B open-weight LLMs, and LoRA fine-tuning the 70B model only matches the small model rather than beating it.
• observationsingle paper↑1↓0
  On human-annotated step-level verification of frontier-model proofs to recent IMO and Putnam problems, most of 29 generative critics and process reward models do little better than chance at finding the first wrong step, a leading math PRM (Qwen2.5-Math-PRM-72B) reaches only about 37% balanced F1, and beyond a few standouts open-source verifiers lag closed models; verification is harder than generation for the same frontier models.
• observationsingle paper↑1↓0
  Outside mathematics, the step-level error detection of both process reward models and general LLMs is markedly weaker than on math benchmarks, and the two fail differently: PRMs miss knowledge-based errors while LLMs miss computational errors, across science and logic domains.
• mechanismsingle paper↑1↓0
  Reward-related information in a reasoning model's hidden states is concentrated in a sparse set of neurons: a linear probe trained with a temporal-difference objective finds 'value neurons' (under 1% per layer) whose activations predict expected correctness, and zeroing them in one layer drops MATH500 accuracy from 75% to an average 20% (as low as 1%), while zeroing the same number of random or magnitude-selected neurons costs 1-9 points; the step-to-step change in the probed value ('dopamine neurons', a stated analogy) used as a process reward for inference-time search lifted MATH500 from 72.2 to 77.8, above an implicit-PRM baseline at 75.0.
• mechanismsingle paper↑1↓0
  For step-level verification of structured quantitative STEM reasoning, training a 7B process reward model only on hard negatives that are constructed to pass a deterministic symbolic verifier (swapped formula or unit-matched variable, with values recomputed so the math balances), and running the verifier as a hard filter before the PRM scores candidates, detects flawed steps better than monolithic PRMs that must judge arithmetic and logic together.
• mechanismsingle paper↑1↓0
  When training an 8B multimodal model with RL to find scientific errors in whole papers without being told where to look, mixing in matched easier variants of the same instances — where the error is asserted and the evidence supplied — yields better performance on the evidence-absent task itself than training only on that task at matched exposure.
• observationsingle paper↑1↓0
  Zero-shot, at function level, small locally run models (Llama 3.2, Mistral) identify 43-45% of 349 real Python bugs from BugsInPy, well above a sub-5% keyword baseline, but with accuracy ranging from 0% to 100% by project and lowest (about 22%) on bugs that span several components, so detection is real but shallow and codebase-dependent.

Techniques

• Pre-train a compact encoder on in-domain reports for consisttraining
  Train a small BERT from scratch on a hospital's own radiology report corpus with masked language modelling plus a findings-conclusion consistency objective, then fine-tune it to classify synthetic reporting errors.
• Probe-derived value signal as a free process rewardtooling
  Fit a linear TD probe on hidden states to get a per-step value estimate, take its step-to-step change as a reward prediction error, and use that as the scoring signal for best-of-n or search at inference, so no separate reward model is run.
• Train evidence-absent verification through matched easier tatraining
  Rewrite each annotated paper error into a chain of three queries — error asserted with evidence given, candidate evidence only, and neither — and train on all three stages jointly rather than only on the hardest one.
• Train the step verifier only on symbolically-valid wrong stetraining
  Generate hard negatives by swapping the logical principle or a unit-matched operand and recomputing values with a solver so they pass a deterministic unit/arithmetic verifier, train the PRM on those pairs, and at inference prune with the ve
• Verify candidates from step-boundary hidden states, not textarchitecture
  Extract the frozen generator's last-layer hidden states at reasoning-step delimiters, feed the step sequence to a 2-layer Transformer encoder, and score candidates with a tie-safe pairwise ranking loss trained on self-sampled trajectories w

Sources

Every paper, post or observation behind this capability. Ones cited by a claim show the claim and its stance; ones filed directly are about the capability but have no claim drafted from them yet. Open one to add support or contest it.

• HSRM: Hidden-State Reward Models for Test-Time Verification2026-08-31
  ↑ supports a claim here
• Error Detection for PET/CT Radiology Reports: Domain-Specific vs Large Language Models2026-08-30
  ↑ supports a claim here
• Not Just Reason, Not Just Scan: Reinforcement Learning for Proactive Scientific Error Verification over Academic Paper2026-08-27
  ↑ supports a claim here
• Neuro-symbolic PRM: Enhancing Scientific Reasoning via Structured Traces and Symbolic Verification2026-08-26
  ↑ supports a claim here
• GR-Ben: A General Reasoning Benchmark for Evaluating Process Reward Models2026-05-02
  ↑ supports a claim here
• An Empirical Evaluation of Locally Deployed LLMs for Bug Detection in Python Code2026-04-25
  ↑ supports a claim here
• Sparse Reward Subsystem in Large Language Models2026-02-01
  ↑ supports a claim here
• Hard2Verify: A Step-Level Verification Benchmark for Open-Ended Frontier Math2025-10-15
  ↑ supports a claim here