FEEDBACK FRICTION: LLMs Struggle to Fully Incorporate External Feedback

We thank the reviewer for recognizing that our paper is well explained with a clear experimental setup, that quantifying receptiveness to feedback is important research, and that our observation and mitigation of this phenomenon represents a key contribution. We have tested additional reasoning models that confirm RIGID THINKING persists, clarified our model-specific temperature settings, and conducted multi-turn interaction experiments. Below we provide detailed responses to your specific concerns:

---

The scope of models tested is limited to LLaMA family only, no reasoning model is tested

TL;DR: We tested Claude 3.7 reasoning models and found the same RIGID THINKING behavior persists.

Although we did not evaluate reasoning models in our NeurIPS submission, we have since tested larger models like Claude 3.7 Sonnet (both with and without Extended Thinking). Table 1 shows the initial accuracy, final accuracy after 10 feedback iterations, and theoretical target accuracy for each model. Despite Claude models achieving significantly higher performance than Llama models, on several datasets, they still plateau well below their target accuracy on several datasets, demonstrating that RIGID THINKING persists across model scales. We'd definitely welcome suggestions for additional models to test.

Table 1: Performance comparison of Claude 3.7 models showing initial accuracy, final accuracy after 10 feedback iterations, and theoretical target accuracy across different datasets.

---

Temperature setting (temperature of 0 is chosen throughout the analysis, even for Qwen models)

TL;DR: In the current submission, we used appropriate temperatures for each model, not uniformly temperature=0.

We want to clarify our temperature choices, which were not uniformly set to 0 as the reviewer suggests:

We tested Qwen3-235B-A22B with temperature=0.7 (following the suggested temperature from their documentation), but found that when reasoning mode is enabled, it still enters repetitive generation loops frequently, and it's not very good at following instructions (the community observed a similar phenomenon), which led to our final decision not to use it.

In addition, in our run with Claude 3.7 with Extended Thinking, we use temperature=1 as suggested by Anthropic.

We choose temperature=0 for most of our testing to ensure deterministic results crucial for controlled experimentation. However, we did conduct experiments with higher temperatures in later sections (§4.2), where we explore progressive temperature increases and rejection sampling as potential mitigation strategies.

---

Human-led study into factors causing Rigid Thinking would be difficult to scale

Yes, as you pointed out, that is a major roadblock if we were to only rely on human evaluation to under Rigid Thinking! However, we did this time-consuming human evaluation for the exact opposite purpose — to verify that o4-mini's judgment of why LLMs didn't follow instructions is correct, so that we don't need to do more human annotation ourselves later.

---

No clear argument for why the effect exists or root cause analysis. Confidence measurement insufficiently justified. Should test on harder tasks (AIME) for stronger signal

TL;DR: We acknowledge we cannot fully explain the root cause, but our confidence analysis on AIME confirms no correlation.

You raised an excellent point! We acknowledge that we cannot definitively explain why this effect exists. We believe a better understanding likely involves complex interactions between how models understand feedback, follow instructions, and update beliefs, which we continue investigating.

Regarding our confidence measurement, we adopted the approach from ClashEval: Quantifying the tug-of-war between an llm's internal prior and external evidence where it demonstrated effectiveness. Since submission, colleagues have suggested alternative confidence/uncertainty measures like semantic entropy (Detecting hallucinations in large language models using semantic entropy ), which we are currently exploring.

We conducted the same confidence experiments on AIME as you suggested. The results show the same conclusion as our other confidence experiments: there is no strong correlation between confidence measures and resistance to feedback.

---

How would using conversation-based feedback (repeated user-assistant iterations), instead of a large prompt containing all the iterations, change the results?

This is a very good suggestion! We have tried to reframe the conversation history into a multi-turn setting by inserting <user> and <assistant> tags for different models and found the performance to be basically the same - LLMs still couldn't reach perfect accuracy with that approach. However, we acknowledge this could be explored more systematically as future work.

---

Can the authors address the comments about their confidence comparison I made in Strengths and Weaknesses?

We have already addressed this in previous comments.

---

Why is a temperature of 0 necessary for quantifying the phenomenon, and why are different temperature schedules not considered in the main evaluation?

We would be more than happy to test other temperatures that you think are worth trying.

As for the temperature setting in this submission, we choose temperature=0 for most of our testing to ensure deterministic results crucial for controlled experimentation. However, we did conduct experiments with higher temperatures in later sections (§4.2), where we explore progressive temperature increases and rejection sampling as potential mitigation strategies.

---

Why did the authors not consider scaling using automated tools?

In fact, we did scale using automated tools (specifically, o4-mini) to categorize the reasons why LLMs fail to integrate feedback. The human annotation is only to test whether this automatic tool is correct or not.

---

Nitpick: masking the boxed answer is often insufficient.

That's a very valid suggestion! We have thought about it and couldn't really find a way to do more masking. For our synthetic tasks (multiplication and hex multiplication), we did mask more intermediate steps, but for other tasks, it's hard to really mask intermediate steps. Also, the fact that LLMs still plateau under these good conditions actually enhances our argument.

---

Rating: 2: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed ethical considerations.

We would sincerely appreciate further elaboration, as we are confident that any such concerns can be addressed in this rebuttal period.

I thank the authors for their thorough rebuttal and further experimental results! I believe most of my concerns with the paper have been addressed, and follow-up on some of the points raised below.

We tested Claude 3.7 reasoning models and found the same RIGID THINKING behavior persists.

These are some exciting results! It is interesting to see this behaviour also appears in models that are already used in production, especially on the level of Claude 3.7. I believe it would be valuable to the LLM reasoning community if the authors consider turning this into a more thorough benchmark. Being able to have a proxy for judging instruction-following capabilities can be inherently useful to determine the applicability of certain models in agentic or RAG-based systems.

Since submission, colleagues have suggested alternative confidence/uncertainty measures like semantic entropy (Detecting hallucinations in large language models using semantic entropy ), which we are currently exploring.

If you manage to reproduce your results with different metrics, please let me know. I believe that proving that your measurement is metric-agnostic will significantly strengthen the claim.

We conducted the same confidence experiments on AIME as you suggested. The results show the same conclusion as our other confidence experiments

Could the authors present their numerical results for this experiment, as well as the conversation-based feedback, in a tabular format?

---

Overall Given the high quality of the rebuttal, and granted the authors include their results in the final revision of their paper, I will raise my score.

We thank the reviewer for recognizing the breadth of our benchmark evaluation, the thoroughness of our error analysis, and the clarity of our presentation. We have conducted additional experiments with reasoning models (Claude 3.7) that confirm RIGID THINKING persists across model architectures, clarified our temperature settings which were model-appropriate rather than uniformly zero, and tested multi-turn interactions which show the same plateauing behavior. Below we provide detailed responses to your specific concerns:

---

Model selection (no proprietary reasoning models has been tested in this submission)

TL;DR: We tested Claude 3.7 reasoning models and found the same RIGID THINKING behavior persists.

Although we did not evaluate reasoning models in our NeurIPS submission, we have since tested proprietary reasoning models such as Claude 3.7 Sonnet and Claude 3.7 Sonnet with Extended Thinking (on‑par with o4‑mini in reasoning capabilities). Table 1 reports shows the initial accuracy, final accuracy after 10 feedback iterations, and theoretical target accuracy for each model. Importantly, even these sophisticated reasoning models exhibit the same fundamental plateauing behavior we term RIGID THINKING, suggesting this limitation persists across different model architectures and reasoning capabilities.

Table 1: Performance comparison of Claude 3.7 models across all nine datasets, sorted by average performance gap per dataset. Despite achieving high final accuracies, models consistently plateau below their theoretical targets across all tasks.

---

Temperature setting (the temperature of 0 is chosen throughout the analysis, even for Qwen models)

TL;DR: In the current submission, we used appropriate temperatures for each model, not uniformly temperature=0.

We want to clarify our temperature choices, which were not uniformly set to 0 as the reviewer suggests: We tested Qwen3-235B-A22B with temperature=0.7 (following the suggested temperature from their documentation), but found that when reasoning mode is enabled, it still enters repetitive generation loops a lot, and it’s not very good at following instructions (community also observed a similar phenomenon), which leads to our final decision to not use it.

In addition, in our run with Claude 3.7 with Extended Thinking, we use temperature=1 as suggested by Anthropic. We choose temperature=0 for most of our testing to ensure deterministic results crucial for controlled experimentation. However, we did conduct experiments with higher temperatures in later sections (§4.2), where we explore progressive temperature increases and rejection sampling as potential mitigation strategies.

---

Prompt structure (can we make the solver/feedback generation into a multi-turn setting)

TL;DR: Multi-turn interactions show about the same performance as single-prompt approach.

Turning feedback into a multi-turn conversational style is a very good suggestion! We have tried to reframe the conversation history into a multi-turn setting by inserting <user> and <assistant> tags, and found the performance to be basically the same - LLMs still couldn't reach perfect accuracy with that approach. However, we acknowledge this could be explored more systematically as future work.

---

Feedback correctness (does incorrect feedback only occur for low-accuracy tasks? What will happen if the feedback is correct in some settings)

TL;DR: Incorrect feedback occurs at all accuracy levels, and better feedback wouldn't solve the core problem.

Regarding whether incorrect feedback is primarily an issue for low or high performance ranges: We have looked into the logs, and this is an issue for all accuracy levels, not just low performance ranges. For the specific case of Llama-4-Maverick on GPQA with 85.7% feedback resistance: If the remaining 5% of incorrect cases received better feedback, the error rate would be approximately 5% × 0.857 ≈ 4.3%, which is still below perfect performance. Also, it's important to note that the "feedback quality" issues don't necessarily mean the feedback is objectively wrong - rather, it's more that the feedback is not targeted enough to address the specific errors made by that model instance.

---

Add baseline comparison in Fig. 5

Thanks for your suggestion! We will add a "no change" baseline line for clearer comparison.

---

Rating: 2: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed ethical considerations.

We would sincerely appreciate further elaboration, as we are confident that any such concerns can be addressed in this rebuttal period.

Hi,

I have no new questions for the authors. It will properly look through all the feedback provided (especially in the discussion with Reviewer xvok) and update my opinion and scores accordingly. Thanks!

We thank the reviewer for recognizing that our paper addresses a crucial and underexplored limitation with rigorous experimental design, and that the concept of RIGID THINKING is clearly defined and well-supported. We have conducted pilot training experiments on smaller models, tested larger models that confirm the phenomenon persists across scales, and explored diverse sampling approaches. Below we provide detailed responses to your specific concerns:

---

The paper does not deeply connect the observed limitations with potential architectural or training modifications, beyond simple sampling tricks.

TL;DR: We conducted pilot training experiments (SFT/DPO) on smaller models but found no improvement in feedback resistance.

In the current submission, we did not perform training interventions such as SFT or DPO on our target models due to resource constraints—the smallest model, LLaMA‑3.3‑70B, requires lots of GPU‑hours for a single fine‑tuning run, which is infeasible within our compute budget.

To verify whether these interventions could improve feedback integration, we conducted pilot experiments on a smaller model, LLaMA‑3.1‑8B‑Instruct, using:

• SFT: 2 000 successful correction trajectories
• DPO: 4 000 preference pairs (2 000 positive / 2 000 negative)
• SFT steps: 5 000 updates, lr = 1 × 10⁻6, batch = 4
• DPO steps: 10 000 updates, same lr & batch
• Evaluation: Feedback Resistance (FR) after 10 feedback iterations on MATH dev

Table 1: Impact of SFT and DPO on Feedback Resistance in LLaMA‑3.1‑8B‑Instruct.

These interventions didn’t result in any FR reduction, we hypothesis this might be related to the findings from Kumar et al. (2024) (Training Language Models to Self-Correct via Reinforcement Learning), who demonstrate that offline SFT methods suffer from distribution shift and behavior collapse, limiting their ability to improve self‑correction. We will keep exploring more integrated training paradigms—such as multi‑turn reinforcement learning that incorporates feedback signals in real time—to enable sustained iterative improvement beyond simple fine‑tuning.

---

Some of the tasks (e.g., synthetic multiplication) are very narrow, so broader generalization to more open-ended and subjective tasks (e.g. summarization) remains uncertain.

Thanks for this feedback! We tested 9 different tasks to capture diverse reasoning patterns, and we'd welcome suggestions for additional tasks you think would be valuable. Regarding more open-ended and subjective tasks like summarization, as we noted in line 124, we deliberately chose objective tasks to avoid reward hacking and reliability issues during evaluation. If you have ideas for how to reliably evaluate feedback incorporation in subjective tasks without these issues, we'd be very interested to hear them.

---

Is there any correlation between model size and resistance to feedback — would even larger models behave differently?

TL;DR: We tested larger models (Claude 3.7) and found RIGID THINKING persists across model scales.

While this NeurIPS submission focuses on the Llama family, we have since tested larger models like Claude 3.7 Sonnet (both with and without Extended Thinking). Table 2 shows the initial accuracy, final accuracy after 10 feedback iterations, and theoretical target accuracy for each model. Despite Claude models achieving significantly higher performance than Llama models, on several datasets, they still plateau well below their target accuracy, demonstrating that RIGID THINKING persists across model scales. We'd definitely welcome suggestions for additional models to test.

Table 2: Performance comparison of Claude 3.7 models across all nine datasets, sorted by average performance gap per dataset. Despite achieving high final accuracies, models consistently plateau below their theoretical targets across all tasks.

---

The study is limited to instruction-tuned LLMs with deterministic decoding setups, which may not reflect real-world use cases involving more diverse sampling.

TL;DR: We used deterministic decoding for reproducibility but tested diverse sampling in §4.2 - the limitation persists.

You're absolutely right about real-world usage patterns. We'd be happy to test other temperature settings you think would be particularly informative. Regarding our temperature choices: we used temperature=0 for most experiments to ensure deterministic, reproducible results crucial for controlled experimentation. However, we did conduct experiments with higher temperatures in §4.2, exploring progressive temperature increases and rejection sampling as potential mitigation strategies. The fact that even these more diverse sampling approaches couldn't fully overcome the plateau suggests the limitation runs deeper than just deterministic decoding.

We thank the reviewer for recognizing that our paper discusses feedback resistance as a model-side limitation rather than a feedback quality issue, and that we show improvement with increased temperature and rejection sampling. We have tested larger models that confirm RIGID THINKING persists across scales and experimented with various prompt variations. Below we provide detailed responses to your specific concerns:

---

There are existing works that show models don't respond to feedback with hypotheses like model size. Comparison with existing work analysis is missing.

Thank you for pointing this out. We would greatly appreciate it if you could share references to those papers, as we'd love to read them and properly compare our findings with prior work. Understanding how our results relate to existing hypotheses about model size and feedback resistance would strengthen our analysis.

---

More experiments are needed for drawing conclusions: need to test more models with different model sizes. Prompt variations are also needed to be tested (e.g. clarifying feedback with examples)

TL;DR: We tested larger models (Claude 3.7) and various prompt variations - RIGID THINKING persists in both cases.

You're absolutely right about testing more models. While this NeurIPS submission focuses on the Llama family, we have since tested larger models like Claude 3.7 Sonnet (both with and without Extended Thinking). Table 2 shows the initial accuracy, final accuracy after 10 feedback iterations, and theoretical target accuracy for each model. Despite Claude models achieving significantly higher performance than Llama models, they still plateau well below their target accuracy, demonstrating that RIGID THINKING persists across model scales. We'd definitely welcome suggestions for additional models to test.

Table 1: Performance comparison of Claude 3.7 models across all nine datasets, sorted by average performance gap per dataset. Despite achieving high final accuracies, models consistently plateau below their theoretical targets across all tasks.

Regarding prompt variations, we experimented with several approaches: we tested more explicit feedback prompts (e.g., "Your error is in step 3 where you incorrectly calculated X"), step-by-step guidance prompts, and example-based feedback where we provided similar problem-solution pairs alongside the correction. Following your suggestion about clarifying feedback with examples, we also tested providing worked examples of similar problems within the feedback itself. Unfortunately, across all these variations, we observed the same fundamental pattern - models still plateau well below their theoretical maximum accuracy.

---

The paper should compare the results against the baseline where feedback doesn't come from the ground-truth answers.

That's an interesting suggestion. Our focus was on establishing an upper bound scenario - if models can't effectively use near-perfect feedback, it suggests fundamental limitations in feedback incorporation. Could you elaborate on what insights you think a weaker feedback baseline would provide? We're curious about your reasoning and we’re more than happy to conduct new experiments, as it might reveal additional aspects of this phenomenon we haven't considered.