Bridging the Know-Act Gap via Task-Level Autoregressive Reasoning

Jihyun Janice Ahn; Ryo Kamoi; Berk Atil; Renze Lou; WonWoo Kang; Heehyun Park; Sarkar Snigdha Sarathi Das; Zhuoyang Zou

cs.AIMar 23, 2026

Bridging the Know-Act Gap via Task-Level Autoregressive Reasoning

Authors: Jihyun Janice Ahn, Ryo Kamoi, Berk Atil, Renze Lou, WonWoo Kang, Heehyun Park, Sarkar Snigdha Sarathi Das, Zhuoyang Zou, Xiaoxin Lu, Yusen Zhang, Asfahan Shah, Ridwanul Hasan Tanvir, Lingxiao Zhao, Hongxi Huang, Vignesh Venkatesh, Dianjun Lin, Hamid Shah, Wentao Wang, Zhanpeng Song, Joshua Reed Bassin, Dax Patel, Ishan Appareddy Agrahar, Sahil Pardasani, Xin Dong, Fatemeh Rahbari, Benjamin David Rishel, Soochan Andrew Lee, Yuv Boghani, Ali B. AlNaseeb, Pranav Suby, Seokhyeon Bae, Shreya Buddharaju, Damien Kula, Soumyadeep Das, Hanyang Frank Liu, Faye Mo, Wenpeng Yin

Organizations: Penn State · UIUC · Columbia University · Episcopal Academy

Abstract

LLMs often generate seemingly valid answers to flawed or ill-posed inputs. This is not due to missing knowledge: under discriminative prompting, the same models can mostly identify such issues, yet fail to reflect this in standard generative responses. This reveals a fundamental know-act gap between discriminative recognition and generative behavior. Prior work largely characterizes this issue in narrow settings, such as math word problems or question answering, with limited focus on how to integrate these two modes. In this work, we present a comprehensive analysis using FaultyScience, a newly constructed large-scale, cross-disciplinary benchmark of faulty scientific questions. We show that the gap is pervasive and stems from token-level autoregression, which entangles task selection (validate vs. answer) with content generation, preventing discriminative knowledge from being utilized. To address this, we propose DeIllusionLLM, a task-level autoregressive framework that explicitly models this decision. Through self-distillation, the model unifies discriminative judgment and generative reasoning within a single backbone. Empirically, DeIllusionLLM substantially reduces answer-despite-error failures under natural prompting while maintaining general reasoning performance, demonstrating that self-distillation is an effective and scalable solution for bridging the discriminative-generative know-act gap

Figures & tables

Question	Balanced binary data: accuracy 95%, prec. 90%, recall 85%, compute F1.
Ask GPT4 to respond (w/o hint)	$F1=\frac{2PR}{P+R}=\frac{2\times 0.90\times 0.85}{0.90+0.85}=0.874$
Ask GPT4 if the question is problematic	On a balanced dataset, these metrics (accuracy = 95%, precision = 90%, recall = 85%) cannot all hold simultaneously. At least one is invalid, so any F1 derived is unreliable.

Table 1: Example where GPT4 fails in generative mode without hints but recognize the issues in discriminative mode.

	Count	Ratio (%)
Discipline
Physical Science	3014	19.07
Biological Science	2776	17.57
Earth Science	2404	15.21
Mathematics	2129	13.47
Social Science	2070	13.10

Table 2: Statistics of FaultyScience .

Model	$P^{Gen}_{-hint}$	$P^{Gen}_{+hint}$	$P^{Dis}$
Mixtral-8x7B	10.2	27.8	75.2
Llama3.3 70B	14.0	24.5	90.8
Qwen2.5-72B	10.8	40.1	95.0
GPT4	34.9	48.6	88.1

Table 3: Answer to

\mathcal{Q}_{1.1}

–performance (%) of existing LLMs in recognizing the issues in the questions of FaultyScience under different prompt style.

Figure 1: Overview of the proposed DeIllusionLLM method.

Model	FaultyScience ( $P^{Gen}_{-hint}$ )	BBH	GPQA
Mixtral-8x7B	10.2	16.58	30.13
Llama3.3 70B	14.0	17.11	21.88
Qwen2.5-72B	10.8	42.78	34.15
DeIllusionLLM (self-distilled from Qwen)	67.8	41.71	33.93
GPT-4	34.9	69.50	34.20
DeIllusionLLM (distilled from GPT)	42.7	68.98	38.80

Table 4: DeIllusionLLM vs. baseline LLMs in three benchmarks.

Model	Earth	Physics	Social	Math	Bio	Eng.	Chem.	CS	mean
GPT-4	37.6	18.4	56.0	33.6	31.2	58.4	24.0	7.2	33.3
DeIllusionLLM	60.8	19.2	56.8	30.4	46.4	72.0	22.4	8.8	39.6

Table 5: Model performance, under

P^{Gen}_{-hint}

, on different disciplines of FaultyScience .

Figure 2: Mode [D] percentages for each dataset.

Aligned (mode = response)
Category	FaultyScience ( $P^{Gen}_{-hint}$ )	BBH	GPQA
[D] + response_D	51.70%	5.35%	0.00%
[G] + response_G	20.80%	60.43%	80.13%
Misaligned (mode $\neq$ response)
[D] + response_G	2.40%	26.74%	19.87%
[G] + response_D	25.10%	7.48%	0.00%

Table 6: Mode–response alignment distribution across datasets.

Figure 3: Ablation study.

Scientific Reasoning in Language Models

Scientific reasoning in language models studies how AI systems use evidence, domain knowledge, and tools to solve problems and support discovery.

See all 133 papers

Answer-Conditioned Chains of Thought Degrade Verifiable-Reasoning Distillation in Large Language Models
Jul 16, 2026Jungseob Lee, Seungyoon Lee, Suhyune Son +4CoT Distillation Language Model Distillation
1Korea University · 2Zoom Communications · 3Yonsei University
Evaluating and Improving LLM Self-Modeling
Aug 31, 2026Siqi Zeng, Andre N. Assis, Rowan WangLLM Evaluation Language Model Introspection
University of Illinois · Constellation · Anthropic
Can LLMs Model Incorrect Student Reasoning? A Case Study on Distractor Generation
Mar 16, 2026Yanick Zengaffinen, Andreas Opedal, Donya Rooein +3LLM Prompting Generative AI in Education
1ETH Zürich · 2Bocconi University · University of Central Florida

Jul 16, 2026J/K move · Enter open · S save

Answer-Conditioned Chains of Thought Degrade Verifiable-Reasoning Distillation in Large Language Models

Jungseob Lee, Seungyoon Lee, Suhyune Son, Dongyub Jude Lee+3

1Korea University · 2Zoom Communications · 3Yonsei University

CoT Distillation Language Model Distillation LLM Reasoning Reasoning Verification Structured Reasoning in Language Models

Open paper AskPDF arXiv

A standard recipe for distilling the reasoning ability of large language models (LLMs) is to sample chains of thought from the model, keep those that reach the correct final answer, and fine-tune on the survivors. When sampling fails, a common fix shows the generator the gold answer and asks it to write a chain that reaches that answer. We show that this second step degrades the training data in a way that correctness filtering cannot catch. We run a controlled experiment that fixes the generator, the problem set, and the correctness filter, and varies only whether the chain is generated under answer-conditioning, the gold answer shown with a request to reach it. Training a strong instruction-tuned reasoning model on its own answer-conditioned chains sharply lowers its verifiable-reasoning accuracy. The loss grows with difficulty, reaching as much as about 27 points on the hardest competition problems. The mechanism is legible in the chains themselves, which rationalize backward from the shown answer instead of deriving it, with the early final-answer statement as the measurable symptom. The harm is a property of the data rather than the generator, read off unlabeled generations before any fine-tuning, ordering the penalty across eight thinking models from four families, and transferring across teacher families. A prompt ablation localizes it to the rationalize-toward instruction rather than the answer's bare visibility. The practical takeaway is to generate answer-blind, because no correctness filter can see this damage in the data.

Abstract

Scientific Reasoning in Language Models

Explore similar work

Answer-Conditioned Chains of Thought Degrade Verifiable-Reasoning Distillation in Large Language Models

Evaluating and Improving LLM Self-Modeling

Can LLMs Model Incorrect Student Reasoning? A Case Study on Distractor Generation