Abstract
Large language models (LLMs) are often deemed unsafe for clinical question answering because of their tendency to hallucinate. Retrieval augmentation, fine-tuning, and external verifiers require new infrastructure that clinical governance must approve and may add latency or extra model calls. Inference-time correction uses the model's internal logit signals, but a fixed transformation need not suit every question. A corrector that improves accuracy by about ten percentage points on a truthfulness stress test yields negligible gains on clinical multiple-choice benchmarks, where instruction tuning concentrates output probability on one answer and leaves low terminal entropy. We introduce ALTAS, which reads terminal entropy and late-layer linearity (R2) from one forward pass to choose per question between greedy decoding and late-layer trajectory correction. No classifier, probe, or head is trained; the router operates on candidate-answer logits and adds 6.5% latency overhead. Applied to every question, the correction improves TruthfulQA over greedy at 3B and 8B by 11.4 and 10.0 percentage points, respectively (p<10−10). Gated per question, ALTAS retains gains of 8.3 to 9.5 percentage points while keeping MedQA, PubMedQA, and MedHallu within a one-percentage-point do-no-harm band, with no statistically significant differences from greedy. The method passes verification sweeps over frozen thresholds, the scoring rule, and the domain label.
Explore similar work
May 7, 2026cs.LG
LLMs reliably correct false claims when presented in isolation, yet when the same claims are embedded in task-oriented requests, they often comply rather than correct. We term this failure mode \emph{correction suppression} and construct a benchmark of 300 false premises to systematically evaluate it across eight models. Suppression rates range from 19% to 90%, with four models exceeding 80%, establishing correction suppression as a prevalent and severe phenomenon. Mechanistic analysis reveals that suppression is not a knowledge failure: the model registers the error internally but task context diverts early-layer attention from the false claim as output intent crystallizes toward compliance at middle layers. We characterize this as \emph{knowing but not correcting} -- suppression occurs at response selection rather than knowledge encoding. Guided by this mechanism, we propose two training-free interventions. Correction Direction Steering (CDS) estimates a correction-compliance direction from matched pairs and injects it at middle layers before output intent crystallizes. Dynamic Payload Amplification (DPA) localizes payload tokens via attention divergence between early and late layers and amplifies their representation at the final layer, requiring no calibration data. Experiments on Qwen3.5-9B and LLaMA3.1-8B show both methods substantially improve factual strictness. CDS achieves the highest correction rate on Qwen3.5-9B (0%
→58.2%). DPA is the only method that preserves or improves reasoning capability on both models. These findings introduce \emph{factual strictness} -- the willingness to uphold accuracy against contextual pressures -- as a new dimension of model reliability.
Zixuan Chen, Hao Lin, Zizhe Chen +6
May 7, 2026cs.AI
Can linearly decodable failure signals in LLM hidden states be leveraged to correct those failures? We investigate this classification-correction gap via Overthinking (OT)--a stable behavioral regime (Jaccard >= 0.81, 94% inter-annotator agreement) in medical QA where models answer correctly under resampling yet fail in extended chain-of-thought. OT is linearly decodable at 71.6% balanced accuracy (p < 10^{-16}). Yet five families of fixed linear steering (29 configurations, n=1,273) all yield Delta ~= 0, with identical null results cross-architecture (Qwen2.5-7B) and cross-domain (MMLU-STEM). Three convergent lines of evidence suggest representational entanglement: the OT direction has 85-88% overlap with task-critical computation (specificity ratio <= 0.152); non-targeted shared-direction steering damages accuracy (-12.1pp); and LEACE concept erasure damages accuracy (-3.6pp, p=0.01), while 10 random erasures produce Delta=+0.3pp. The per-instance probe-steering correlation is r=-0.002 (p=0.97). Positively, the same probe enables selective abstention (held-out AUROC=0.610, exceeding all five uncertainty baselines, p=0.009): decodable failure structure supports post-generation reliability estimation even when the fixed linear steering family cannot exploit it for correction.
Ming Liu
Jun 25, 2026stat.ML
Many decoding methods for large language models can be understood as shifting probability mass toward outputs that are more likely under the model, either locally at the token level or globally at the sequence level. Therefore, their success depends on a fundamental question: when does sequence probability, that is, the conditional probability of a continuation given a prompt, actually align with correctness? In this paper, we set out to quantify this relationship across decoding methods, models, and benchmarks at four levels: across decoding methods, across hyperparameters within a method, across prompt-answer pairs within a dataset, and across repeated responses to the same prompt. We find that higher sequence probability is often predictive of correctness across prompt-answer pairs within a fixed dataset. However, this relationship does not generally transfer to decoding decisions: increasing sequence probability by changing hyperparameters or methods does not reliably improve accuracy. Further, sequence probability is not a good indicator of correctness for responses to the same prompt. These findings clarify when decoding can and cannot be expected to improve correctness, and provide practical guidance for decoding, self-consistency, and verifier-free self-improvement.
Johannes Zenn, Jonas Geiping