Knowing When Not to Answer: Abstention-Aware Scientific Reasoning

1. Problem Statement

Large Language Models (LLMs) are increasingly deployed to verify scientific claims and answer biomedical questions. However, current evaluation paradigms assume models must always produce a definitive answer (unconditional accuracy). In scientific contexts, this is problematic because:

• Harm of Error: An incorrect conclusion (e.g., endorsing an ineffective treatment) is often more harmful than providing no conclusion.
• Evidence Limitations: Scientific reasoning often involves incomplete, ambiguous, or weak evidence, yet standard metrics incentivize models to "guess" rather than abstain.
• Evaluation Gap: Existing benchmarks (like SciFact and PubMedQA) focus exclusively on accuracy, obscuring a model's ability to recognize when evidence is insufficient.

The paper argues that the primary challenge in scientific reasoning is not selecting the "best" model, but determining when available evidence is sufficient to justify an answer.

2. Methodology

The authors propose an Abstention-Aware Verification Framework that decomposes reasoning into auditable steps and explicitly models the option to abstain. The pipeline consists of five key stages:

A. Problem Formulation (Selective Classification)

The task is framed as Selective Classification, where the output space is extended from labels $\mathcal{Y}$ to $\mathcal{Y} \cup \{\perp\}$ (where $\perp$ denotes abstention). The goal is to minimize Selective Risk (error rate on answered queries) subject to a controllable Coverage constraint (fraction of queries answered).

B. Condition Decomposition

Instead of treating a claim or question as a monolithic unit, the input $x$ is decomposed by a generative LLM into a set of $k$ minimal, atomic conditions $C = \{c_1, \dots, c_k\}$.

• Each condition represents a necessary factual requirement.
• Conditions are marked as Critical ($\beta_i=1$) or non-critical. At least one critical condition must exist.
• This step transforms complex reasoning into independent sub-problems.

C. Evidence Auditing via NLI

Each condition $c_i$ is audited against the evidence set $E$ using a pre-trained Natural Language Inference (NLI) model (DeBERTa-based cross-encoder).

• The NLI model computes probabilities for Entailment ($s_{ent}$) and Contradiction ($s_{con}$).
• Audit Status ($a_i$): Assigned based on thresholds:
  • CON (Contradiction): If $s_{con} \ge \theta_{con}$.
  • SUP (Support): If $s_{ent} \ge \theta_{ent}$ and no contradiction.
  • MIS (Missing): Otherwise (treated as lack of evidence, not weak support).
• Asymmetry: Contradiction takes priority over support to reflect the high cost of false positives in science.

D. Decision Aggregation under Asymmetric Loss

A deterministic function $h$ aggregates audit results into a final decision, governed by an Asymmetric Loss Structure ($\ell_{fs} > \ell_{fr} > \ell_{\perp} = 0$), where False Support is costlier than False Refutation.

• Claim Verification (SciFact):
  • Refutes: If any critical condition is contradicted.
  • Supports: If all critical conditions are supported.
  • NEI (Not Enough Info): Otherwise.
• Question Answering (PubMedQA):
  • No: If any critical condition is contradicted.
  • Yes: If at least one critical condition is supported (and none contradicted).
  • Maybe: Otherwise.

E. Confidence Estimation and Abstention

A confidence score is derived from the audit margins:
$$conf(x, E) = \max_{c_i \in C_{crit}} |s_{ent}^i - s_{con}^i|$$
The system abstains if $conf(x, E) < \tau$, where $\tau$ is a tunable threshold. This allows for post-hoc control of the risk-coverage trade-off without retraining.

F. Evaluation Protocol

Performance is measured via Risk-Coverage Curves and Area Under the Risk-Coverage Curve (AURC).

• Risk: Error rate among non-abstained predictions.
• Coverage: Fraction of instances answered.
• Lower AURC indicates better reliability across coverage levels.

3. Key Contributions

1. Abstention-Aware Pipeline: A novel framework that decomposes scientific inputs, audits conditions via NLI, and enables selective prediction based on evidence sufficiency.
2. Systematic Empirical Study: Evaluation across six diverse models (encoder-decoder, open-weight, proprietary APIs) on SciFact and PubMedQA, demonstrating that risk control via abstention is a dominant factor in reliability, often outweighing architectural differences.
3. Unified Evaluation Framework: Introduction of risk-coverage analysis as a standard for scientific reasoning, arguing that evaluating when a model should not answer is more critical than what it answers.

4. Experimental Results

The study evaluated models including Flan-T5, Llama-3.3, Mistral, DeepSeek, GPT-4o-mini, and GPT-5.2.

• Limited Separation in Accuracy: Under unconditional metrics (forcing an answer for every query), raw accuracy and F1 scores clustered tightly (e.g., SciFact accuracy ~0.39–0.40). Larger or proprietary models did not significantly outperform smaller ones.
• Significant Gains in Selective Risk: When abstention was enabled:
  • Risk Reduction: All models showed a sharp decline in error rates as coverage decreased (e.g., halving risk at moderate coverage).
  • Model Agnosticism: The behavior of abstention was more predictive of reliability than the model architecture. A smaller model with strict abstention often outperformed a larger model forced to answer everything.
  • Instruction Tuning: Instruction-tuned models (like Flan-T5) exhibited more conservative decision-making, achieving lower risk at comparable coverage levels.
• Ablation Studies:
  • No Decomposition: Led to earlier, more conservative abstention but lost diagnostic granularity.
  • No Audit: Resulted in "degenerate" high accuracy (due to trivial heuristics) but zero correlation between confidence and correctness, proving evidence auditing is essential.
  • No Abstention: Resulted in the highest risk, confirming that abstention is the primary mechanism for error control.

5. Significance and Implications

• Paradigm Shift: The paper advocates shifting the evaluation metric from "Can the model answer correctly?" to "Can the model recognize when it should not answer?"
• Safety in Science: By prioritizing the minimization of false positives (false support) through asymmetric loss and abstention, the framework aligns with the high-stakes nature of scientific and medical domains.
• Practical Utility: The framework provides a model-agnostic lens for assessing reliability. It suggests that in scientific workflows, the "best" system is not necessarily the most accurate one, but the one that best calibrates its confidence to the available evidence.
• Future Directions: The authors suggest extending this to multimodal data, domain-specific verifiers, and benchmarks that explicitly reward calibrated abstention.

In summary, the paper demonstrates that principled abstention is the key to reliable scientific reasoning with LLMs, offering a robust method to mitigate the risks of hallucination and overconfidence in critical domains.