TLDR: We evaluated the generalization of the classifier-selective method by training on specific constraint subsets, showing improved generalization when focusing on constraints where reasoning typically fails. Additionally, we explored direct instruction-following fine-tuning (SFT), finding it effective but also potentially harmful to performance on unrelated tasks, highlighting the importance of weight-free mitigation methods.

Thank you for recognizing both the novelty of the phenomenon and the comprehensive nature of our analysis and experiments. We also appreciate your acknowledgment of our mitigation strategies and their effectiveness across diverse models and benchmarks. Meanwhile, we sincerely appreciate the comments and provide our responses below:

1. Generalizability of Classifier-Selective Method:

We understand the concern about the generalizability of the classifier-selective method and agree that it is an important aspect to explore. One challenge is that each sample often contains multiple overlapping constraints. Nonetheless, we try our best to implement the idea of training the classifier on a subset of constraint types and evaluating its generalization to other types. Specifically, the IFEval dataset annotates constraints with higher-level categories including:

• ['detectable_content', 'startend', 'combination', 'keywords', 'change_case', 'punctuation', 'detectable_format', 'language', 'length_constraints'].

We selected five types such that the samples containing these categories account for roughly half of the dataset:

• ['length_constraints', 'combination', 'startend', 'language', 'change_case'].

For each model, we trained the classifier only on this subset and evaluated its performance on the held-out half. Interestingly, we observed even stronger results than those from classifiers trained on a random 50% split: in many cases, the improvements were significantly larger than before, as shown below (improvements are relative to the Reason/CoT model, save as in the tables in paper):

First, these results show that the classifier-selective method does generalize across constraint types in certain aspects. Interestingly, they also suggest that training the classifier specifically on certain constraint types may lead to even better overall performance.

To explain this, we hypothesize that it is because the selected types (e.g., length_constraints, combination) are where CoT is most prone to introducing errors. To support this, we conducted a further quantitative analysis of constraint-type distributions. Below we show the Reason-WIN vs. Reason-LOSE distributions on IFEval for two representative models (detailed distributions for all models/tasks will be included in the Appendix of our paper):

GPT-4o-mini

• Reason WIN: keywords 19.5%, length_constraints 17.1%, change_case 24.4%, combination 2.4%, detectable_format 19.5%, detectable_content 9.8%, punctuation 4.9%, startend 2.4%

• Reason LOSE: combination 35.2%, length_constraints 31.0%, change_case 5.6%, keywords 16.9%, startend 2.8%, detectable_content 2.8%, punctuation 4.2%, detectable_format 1.4%

Claude-3.5-haiku

• Reason WIN: change_case 30.0%, keywords 25.0%, length_constraints 15.0%, detectable_format 20.0%, detectable_content 5.0%, combination 5.0%

• Reason LOSE: combination 43.3%, length_constraints 25.4%, keywords 10.4%, startend 4.5%, change_case 4.5%, detectable_content 1.5%, detectable_format 6.0%, language 3.0%, punctuation 1.5%

For both models, we observe a clear pattern: Reason-WINs are concentrated in lexical or formatting categories (e.g., keywords, change_case, detectable_format), while Reason-LOSEs are dominated by global or compositional constraints (e.g., combination, length_constraints, startend). These “failure-prone” types were exactly those we selected for training the classifier, and moreover, these results match and further support our case study insights in Section 4.1.

2. Instruction-Following Training:

We thank the reviewer for this thoughtful suggestion! Indeed, we are actively exploring additional methods to mitigate the drop in instruction-following performance caused by reasoning/CoT. The primary goal of this paper is to reveal and analyze this degradation, supported by quantitative experiments, case studies, and attention analysis. For mitigation, we focused on simple methods that do not alter model weights, because for methods involving training, which change the model weights, our major concern is how this training may affect performance on other tasks.

We explore this with experiments. Certain benchmark datasets (such as ComplexBench) involve both rule-based evaluation and LLM-as-a-judge, so RL training for instruction-following might be expensive, as LLM-as-a-judge could significantly slow down the process (but we will explore this further in future work). Therefore, we conducted supervised fine-tuning (SFT) experiments using data labeled by the strongest model in our evaluation (Claude 3.7-Sonnet). We selected only samples where the model response received a perfect instruction-following score, and trained using these examples as positive supervision.

We split half of the IFEval dataset for SFT training and evaluated instruction-following performance on the other half after SFT (denoted as “IFEval (SFT)”). We also evaluated performance before and after SFT on other representative tasks: ARC, TruthfulQA, GSM8K, and MinervaMath, which include both general knowledge and reasoning evaluation, to assess any unintended impact. Below are the results:

We observe that SFT can significantly improve instruction-following performance for many models (e.g., a +16.0 improvement for Meta-Llama-3-8B-Instruct). However, it can also negatively impact performance on other tasks (possibly due to catastrophic forgetting on the other tasks after SFT on instruction-following). For instance, Meta-Llama-3-8B-Instruct shows performance drops across all evaluated benchmarks after SFT. This highlights the need for caution when applying weight-updating mitigation strategies.

Meanwhile, we are continuing to explore alternative mitigation strategies that preserve model weights. To share some ideas, some directions include:

• Activation probing as a reasoning gate: A lightweight linear probe trained on mid-layer activations could predict whether reasoning is needed. If reasoning is deemed unnecessary, the model uses the base response; otherwise, it switches to CoT.
• Logit-lens comparison: By running Base and Reason side-by-side, we can use the logit lens to identify positions in the prompt where their token distributions diverge most. Examining these hidden-state differences could reveal the essential source of instruction-following failures.
• Activation steering to mitigate failures: We could precompute steering vectors associated with constraint types (or derived from the hidden-state differences above) and inject them during generation. This may nudge the model toward better instruction adherence, without fine-tuning.

Overall, we deeply appreciate the reviewer’s insightful comments and valuable suggestions. We sincerely hope our clarification and additional experiments above would address the concerns, and respectfully hope the reviewer could even further raise the score.

TLDR: We clarify the validity of our two experimental settings, added new controlled experiments to reinforce our findings, conducted rigorous statistical significance tests to demonstrate robustness, and provided additional evaluations showing reasoning may not only reduce instruction-following accuracy but can also degrade overall response quality.

Thank you for highlighting the novelty and importance of our main finding. We also appreciate your recognition of the depth and rigor of our failure mode analysis and manual case study. Below are our responses to the comments:

1. Design of Setting 1 and Setting 2

First, we would like to clarify our motivation and thoughts behind the design of two experimental settings. Setting 1 (CoT prompting) involves using the same underlying model with and without chain-of-thought prompting. This provides a well-controlled comparison since the model weights are identical, isolating the effect of prompting alone. CoT prompting is also a common technique applied in practice to elicit reasoning. As noted in Section 3.4, Setting 2 (comparing a base model to its Reasoning counterpart) introduces potential confounds: Reason models may undergo additional training stages on new datasets, such as supervised fine-tuning or RL. For example, DeepSeek-V3 and DeepSeek-R1 differ significantly in training scale, with R1 having been further trained on much more data in multiple extra training stages, involving both SFT and RL. Despite these differences, we include both settings to demonstrate the general phenomenon: a consistent drop in instruction-following performance after applying CoT or reasoning.

In general, we agree and acknowledge that each setting has distinct advantages and limitations. To further strengthen the Setting 2 results, we added additional base–reasoning pairs:

• Qwen2.5-1.5B-Instruct vs. Qwen2.5-Math-1.5B
• Qwen2.5-7B-Instruct vs. Qwen2.5-Math-7B

In each case, the Math variant is explicitly fine-tuned for reasoning. Furthermore, we incorporated newly released Qwen3 models, which include both Base and Think modes. This is a better controlled experiment as the think vs. no-think mode only differ in the template and the model weights are the same:

• Qwen3-4B vs. Qwen3-4B-Think
• Qwen3-8B vs. Qwen3-8B-Think
• Qwen3-32B vs. Qwen3-32B-Think

We evaluate all pairs on both IFEval and ComplexBench. As shown in the table below, the general trend persists: reasoning variants tend to underperform their base counterparts in instruction-following, across both benchmarks.

We sincerely appreciate the reviewer’s comment and will incorporate these additional results and corresponding discussion into Section 3.4 of the revised paper.

2. Significance Test

In the evaluation setup, the rule-based Python functions are deterministic. For ComplexBench, which includes both rule-based and LLM-as-a-judge components, we also use temperature 0 for GPT-based evaluation. For the Classifier-Selective mitigation method, since it involves classifier training on a randomly sampled subset of the data, we further examine its robustness by running Monte Carlo cross-validation by repeating random hold-out three times. Below we report the average, standard deviation, improvement over the Reason/CoT model, and p-values from two-sided z-tests:

We also computed the overall p-values across all models using paired t-tests on the improvement scores: 6.3e-05 for IFEval, and 2.1e-06 for ComplexBench. These consistently small p-values, along with the positive improvement scores, indicate that the improvements from the Classifier-Selective method are statistically significant and robust. This further strengthens the effectiveness of our proposed approach.

3. Quality (besides Instruction Following) of Response:

The overall response quality, beyond instruction-following performance, is indeed another important and interesting dimension to evaluate. We agree with the reviewer that reasoning can influence both instruction adherence and general response quality. While reasoning models often excel at tasks like math and code generation, the instructions in IFEval and ComplexBench are more about following details and less creative (e.g. writing an email in a specific format).

In our paper, we focused on how reasoning can degrade instruction-following ability. Here, we add new experiments to assess response quality directly. Specifically, we use LLM-as-a-judge (GPT4o) to compare the responses of Base vs. Reason models. Importantly, we restrict this evaluation to only those examples where both models perfectly satisfy all constraints, so that differences in quality are not due to instruction-following violations.

We find that in many cases (even all models on ComplexBench) the base model achieves a win rate above 50%, indicating that it produces higher-quality responses than the reasoning variant, even when both follow instructions perfectly. This result suggests that reasoning can not only hurt instruction following, but may also degrade general response quality in tasks where step-by-step reasoning is unnecessary.

Overall, we deeply appreciate the reviewer’s insightful comments and valuable suggestions. We sincerely hope our clarification and additional experiments above would address the concerns, and respectfully hope the reviewer could raise the score.

Dear Reviewer EvS7,

Thank you again for taking the time to review our paper. We hope that our rebuttal has addressed your concerns and clarified our contributions. If you find our response helpful in resolving the issues raised, we would be sincerely grateful if you would consider updating your score to reflect your current assessment. If any further questions or uncertainties remain, we would be more than happy to further clarify during the discussion period!

Thank you for addressing my concerns. I'm updating my score accordingly

TLDR: We agree exploring multilingual instruction-following evaluation is valuable future work. We’ve clarified that our benchmarks already include everyday tasks (not just reasoning-heavy scenarios), added quality-based evaluations using LLM judges, and incorporated correlation analysis regarding CoT length. Additionally, we discussed further mechanistic interpretability methods and acknowledged the potential of combining multiple mitigation strategies.

Thank you for the positive and detailed feedback. We’re especially glad that you found our evaluation thorough, the constraint-attention analysis insightful, and the methods practical and applicable. We also really appreciate the insightful comments! Below are our responses:

1. Task on Non-English Languages

We really appreciate this suggestion. We agree it would be interesting to study this phenomenon and apply our approaches to other languages. It is out of the scope of this paper but we will actively explore this direction and include multilingual analysis as part of our future work!

2. Evaluation on Non-Reasoning Tasks

If the reviewer means applying our attention-based analysis to non-reasoning tasks, we believe it can generalize. In our study, attention is used to identify peak signals on important tokens: specifically, tokens corresponding to constraints in instruction-following tasks. We examine how the model’s responses relate to these tokens through their attention scores. This approach is not specific to reasoning and should extend naturally to other tasks by redefining what constitutes “important” tokens.

If the reviewer is referring to our instruction-following evaluations, in fact, here the IFEval and ComplexBench benchmarks are not specifically designed for reasoning; rather, they focus on evaluating instruction-following capabilities of LLMs. Most instructions in these datasets are everyday user requests (e.g., writing emails or formatting text), not reasoning-heavy tasks. If the reviewer means evaluating models on other aspects of performance beyond instruction following, we have also added new experiments to assess response quality. Specifically, we use LLM-as-a-judge (GPT4o) to compare the responses of Base vs. Reason models. We restrict this evaluation to only those examples where both models perfectly satisfy all constraints, so that differences in quality are not due to instruction-following violations. Below are the win rates of Base models:

We find that in many cases (even all models on ComplexBench) the base model achieves a win rate above 50%, indicating that it produces higher-quality responses than the reasoning variant, even when both follow instructions perfectly. This result suggests that reasoning can not only hurt instruction following, but may also degrade response quality in tasks where step-by-step reasoning is unnecessary.

3. More Related Work

These papers are indeed relevant and very interesting, and we are excited to see these related works come out recently. We have included them into the related work section of our paper. We really thank the reviewer for mentioning these papers! The exploration of CoT length is particularly intriguing, and we also conducted some preliminary experiments out of curiosity. In Appendix I, we already examined the correlation between the accuracy difference of the Base and Reason models and the length of the reasoning tokens, and observed no clear correlation. Below, we conduct additional experiments to study the correlation between the accuracy difference and the length difference between the Base and Reason models.

More precisely, we conducted this analysis and report the Pearson correlation between accuracy difference (Acc(Base) – Acc(Reason)) and token length difference (Len(Reason) – Len(Base)) for both IFEval and ComplexBench. Overall, we do not observe strong correlations, except for DeepSeekR1-Distill-Qwen7B on IFEval. We omit p-values for brevity, but note that this case had a p-value of 2e-8, the only one below 5%. In this particular case, it suggests that longer reasoning may be significantly associated with a larger drop in instruction-following performance. We will include these results and further discussion in Appendix I.

4. Other Mechanistic Interpretability Methods

We agree that additional MI tools can illuminate the phenomena in our paper. In fact we are actively exploring some other ideas:

• Activation probing as a reasoning gate. A simple linear probe trained on mid-layer activations might be able to predict whether CoT will help. When the probe signals “reasoning unnecessary,” the model stays in Base mode. Otherwise it switches to CoT, reducing harmful reasoning.
• Logit-lens comparison. Running Base and Reason side-by-side, we can project each layer’s residual stream through the LM head (logit lens) and locate constraint-relevant prompt positions where their vocabulary distributions diverge the most. Then inspecting the hidden-state difference at these hotspots should reveal the essential causes of their different performances.
• Activation steering to mitigate the drop. To further mitigate the drop (i.e. improve instruction following ability), another method we could think of is to perform activation steering (pre-compute the steering vectors for different types of constraints, or they could be the hidden-state difference identified above). Injecting a small multiple of this vector at generation time could nudge outputs back toward instruction compliance without changing model weights.

5. Combination of Multiple Strategies

We agree these methods are relatively orthogonal and could be combined, although with potentially increased complexity. In this paper, our main goal is to systematically reveal and analyze the drop in instruction-following caused by reasoning/CoT through controlled experiments, case studies, and attention analysis. Thus, we focused on straightforward mitigation methods that do not alter model weights, since weight changes may negatively affect performance on other tasks. For instance, we conducted additional experiments exploring the strategy of directly training the model for instruction following. We do SFT on an IFEval subset and found that it improved instruction-following for some models but reduced performance on general and reasoning benchmarks (detailed results can be found in “2. Instruction-Following Training” in our response to Reviewer TQ5w below). Further combining multiple mitigation strategies and optimizing their interactions is an exciting direction that we plan to explore in future work!

Overall, we deeply appreciate the reviewer’s insightful comments and valuable suggestions. We sincerely hope our clarification and additional experiments above would address the concerns, and respectfully hope the reviewer could further raise the score.

TLDR: We clarify our choice of lightweight mitigation methods, incorporate suggested related works, and provide statistical analyses supporting our qualitative findings. We also address concerns about evaluation setups by adding Monte Carlo cross-validation, correlation analyses, and further experiments (e.g., CoT on Reason models). Additionally, we clarify figure interpretations and justify our choice of deterministic evaluation settings.

Thank you for the thoughtful and encouraging feedback. We’re especially grateful that you found our motivation, model-scale experiments, proposed methods, and constraint attention analysis to be clear and meaningful contributions. We also appreciate the insightful comments. Below are our responses:

1. Straightforward Mitigation Methods

We agree that the mitigation methods are straightforward. Our main focus is to reveal the drop in instruction-following caused by reasoning/CoT, supported by experiments, case studies, and attention analysis. For mitigation, we focus on light-weight methods that can be applied at inference time. We have limited API access to those proprietary LLMs and methods that involve changing model weights (e.g. specialized post-training recipes) may affect performance on other tasks. For example, we added SFT experiments on an IFEval subset: while some models improved in instruction-following, performance dropped on other general and reasoning tasks (please see results for “2. Instruction-Following Training” in our response to Reviewer TQ5w.)

2. More Related Works

We appreciate the reviewer for highlighting these relevant papers. [1] shows CoT helps link sparsely co-occurring information, and [2] examines the trade-off between instruction-following and faithfulness. Our study is complementary, and we will add the discussion of both in the Related Work section.

3. Statistical Support for Case Study Summary

We conducted a quantitative analysis of general constraint type distributions across models, comparing cases where reasoning helps vs. hurts. Below, we show WIN/LOSE constraint-type distributions for two models on IFEval. (Detailed distributions for all models/tasks will appear in Appendix.)

GPT-4o-mini

• Reason WIN: keywords 19.5%, length_constraints 17.1%, change_case 24.4%, combination 2.4%, detectable_format 19.5%, detectable_content 9.8%, punctuation 4.9%, startend 2.4%

• Reason LOSE: combination 35.2%, length_constraints 31.0%, change_case 5.6%, keywords 16.9%, startend 2.8%, detectable_content 2.8%, punctuation 4.2%, detectable_format 1.4%

Claude-3.5-Haiku

• Reason WIN: change_case 30.0%, keywords 25.0%, length_constraints 15.0%, detectable_format 20.0%, detectable_content 5.0%, combination 5.0%

• Reason LOSE: combination 43.3%, length_constraints 25.4%, keywords 10.4%, startend 4.5%, change_case 4.5%, detectable_content 1.5%, detectable_format 6.0%, language 3.0%, punctuation 1.5%

We notice that for GPT-4o-mini,

Reasoning Helps ✓: 81 % of the Reason‐wins come from format / lexical categories: detectable-format (19.5 %), change-case (24.4 %), keywords (19.5 %), length-constraints (17.1 %).

Reasoning Hurts ✗: 66 % of the Reason‐losses are classic “over-focus / extra content” errors: combination-repeat-prompt & two-responses (35.2 %), length-constraints over-runs (31.0 %). This cleanly mirrors patterns 3 and 4 in our case study in Section 4.1.

Claude-3.5-Haiku shows similar trends: lexical precision dominates wins, while losses are due to extra text/simple counting errors.”

4. Different Evaluation Sample Size for Method 4

We understand the concern regarding different test set sizes. To address this, we performed Monte Carlo cross-validation by repeating random hold-out three times. We also evaluated the latest Qwen3 models (which have thinking mode). Below, we report the average, standard deviation, improvement scores of the mitigation Method 4, and p-values from the two-sided z-tests:

We also computed the overall p-value from a paired t-test on improvements across all models: 6.3e-05 for IFEval and 2.1e-06 for ComplexBench. These significant improvements support the robustness and effectiveness of Method 4.

5. Explanation of Figure Titles (Figure 1 and 2)

We thank the reviewer for pointing this out. The accuracy evaluation for IFEval and ComplexBench is based on the number of instructions the model follows. The list of “True” and “False” in Figures represents boolean values indicating whether each response satisfies the constraints, and the “score” refers to the total count of “True” values. Thank you for pointing this out and we will clarify this in the paper.

6. Correlation between Accuracy and Token Length Difference

We agree that using token length difference may be a more natural choice. We conducted this analysis and reported the Pearson correlation between accuracy difference (Acc(Base) – Acc(Reason)) and token length difference (Len(Reason) – Len(Base)) for both IFEval and ComplexBench. Overall, we do not observe strong correlations, except for DeepSeekR1-Distill-Qwen7B on IFEval. We omit p-values for brevity, but note that this case had a p-value of 2e-8, the only one below 5%. In this particular case, it suggests that longer reasoning may be associated with a larger drop in instruction-following performance. We will include these results and further discussion in Appendix I.

7. CoT Prompt a Reasoning Model

We appreciate the suggestion. We added experiments applying a CoT prompt to a reasoning model (double thinking, denoted “ThinkCoT”) on three Qwen3 models. We observe that it produces longer responses with two reasoning segments, and native thinking tends to be conversational, while prompted CoT more formal. We observe modest gains of "ThinkCoT" over “Think” in some cases (e.g. 8B/32B on ComplexBench, 4B on IFEval), though still below “Base”. In other cases (e.g., 8B/32B on IFEval), performance drops. We do not see correlation to token length as longer “double-thinking” does not show consistent directional performance change over “Think” mode. In this situation, we suspect double thinking makes the model focus too much on thinking, affecting response quality.

8. Manual Examination

For each model, eight human annotators manually examined the responses of the Base and Reason variants. For each sample, we provided constraint-specific accuracy and asked annotators to write a short “TL;DR” summarizing why the Reason response won or lost. We found many recurring patterns in these summaries, which we categorized to support the insights presented in Section 4.1.

9. Choice of Temperature

For our instruction-following tasks and evaluation, we aim for more deterministic responses and scores. While using a nonzero temperature with multiple trials is an option, different models exhibit varying diversity at the same fixed temperature, and a limited number of trials may not accurately reflect true performance. Thus, we believe using temperature 0 is a more natural and reliable setting.

We deeply appreciate the reviewer’s valuable comments. We hope our clarification and additional evaluations would address certain concerns, and sincerely hope the reviewer could raise the score.

Dear Reviewer Rspa,

Thank you again for taking the time to review our paper. We hope that our rebuttal has addressed your concerns and clarified our contributions. If you find our response helpful in resolving the issues raised, we would be sincerely grateful if you would consider updating your score to reflect your current assessment. If any further questions or uncertainties remain, we would be more than happy to further clarify during the discussion period!

We sincerely thank Reviewer Rspa for the kind reply. We are very glad to hear that our rebuttal helped address most of your concerns, and we truly appreciate your consideration in increasing the score!

In fact, due to the length limit in the rebuttal response, we focused more on providing results to address other questions and kept our discussion in Section "2. More Related Works" brief. Below, we provide a more detailed response for this part:

2.1 Connecting to Reference [1]

Thank you for the suggestion to connect our findings with theoretical accounts of when CoT helps. Our paper analyzes when and why CoT helps/hurts at the constraint level through case studies and a constraint-token attention analysis. Empirically, CoT degrades instruction following across models on IFEval and ComplexBench. Moreover, mechanistically, we find CoT often diverts focus away from simple constraints (e.g. exceeding length limits or inserting extra redundant text), while helping some format/lexical constraints.

We will add an explicit discussion linking these observations to [1]. Their theory shows thinking steps help when tasks require bridging non-local dependencies ("reasoning gaps"). In our setting, many IFEval or ComplexBench constraints are locally checkable (such as token-level counts, formatting), so they rarely require such non-local composition. Therefore, the extra reasoning from CoT can reduce attention to constraint tokens, explaining the observed harm. We currently also have a hypothesis: the more an instruction requires non-local composition (e.g. ComplexBench constraints with type Nested/Chain), the less harmful CoT should be, which is consistent with [1]. We will include a small stratified analysis to further investigate this conjecture, while more detailed exploration is left for future work.

2.2 Adding Citation [2]

We agree and will cite [2]. That paper empirically separates two axes: instruction following and faithfulness to provided context. The authors show a trade-off when fine-tuning for one versus the other under a two-stage protocol.

Our contribution is orthogonal and complementary: we study how explicit CoT affects instruction following ability with rule-verifiable constraints (e.g. word counts, formatting, lexical mentions) without external context. We demonstrated consistent CoT-induced drops and analyzed this phenomenon. As noted in our paper, CoT can enhance some output constraints (formatting/lexical precision), indicating that instruction following ability is multifaceted.

Planned changes

• We will add a paragraph connecting our findings to [1]’s locality/reasoning-gap account and state the non-local-composition hypothesis, with preliminary explorations.
• We will also add a scoped paragraph explicitly distinguishing output-format constraints (our focus: IFEval/ComplexBench) from context adherence/faithfulness (RAG-style tasks in [2]). This positions our findings as complementary to the trade-off observed in [2].
• Moreover, we also plan to include a brief note on applicability: our selective-reasoning methods naturally extends to RAG settings by adding a "context present" feature to the selector (i.e. the classifier). This predicts both whether the CoT would help this instruction and whether the context-faithfulness is at risk, which aligns with the observation in [2] that objectives can conflict. We will include this as a suggested extension in our revised paper.

Finally, we are especially grateful for the insightful comments by Reviewer Rspa. Your feedback has led to valuable improvements in our work. We deeply appreciate your thoughtful evaluation and engagement throughout the discussion process.

References

[1] Prystawski, A., et al. Why think step by step? (theoretical account: CoT helps when tasks exhibit non-local structure/reasoning gaps).

[2] Wu, Z., et al. Dancing in Chains: On Faithfulness and Instruction Following in Alignment. (separates output-constraint following from context faithfulness; shows trade-offs under two-stage fine-tuning).