Letting Tutor Personas Speak Up for LLMs: Learning Steering Vectors from Dialogue via Preference Optimization

Alexander Scarlatos; Andrew Lan; Jaewook Lee; Simon Woodhead

REVIEW 3 major objections 4 minor 35 references

Letting Tutor Personas Speak Up for LLMs: Learning Steering Vectors from Dialogue via Preference Optimization

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A single activation-space direction, scaled per tutor, can shift an LLM tutor from an average style toward a specific human tutor's style, learned from real dialogues rather than persona prompts.

desk verdict A useful, modest extension of BiPO to tutor-persona steering, but the central evaluation is undermined by an undefined baseline generation and a lack of error bars. read the letter →

arxiv 2602.07639 v2 pith:RXYGOS5D submitted 2026-02-07 cs.CL

Jaewook Lee , Alexander Scarlatos , Simon Woodhead , Andrew Lan This is my paper · ORCID

classification cs.CL

keywords activationsteeringvectorspreferenceoptimizationtutorpersonastutoringdialoguesLLMstylecontrol

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that the stylistic differences between human math tutors—how much they scaffold, praise, correct, or rush students—can be captured by one direction in an LLM's internal activation space, learned directly from real tutor-student dialogues rather than from persona prompts. It trains a shared steering vector plus a per-tutor scale, moving the model away from an average-tutor baseline toward the target tutor's utterances. If correct, tutor-specific behavior becomes a dialable, interpretable control rather than a prompt-engineering problem. The authors report improved semantic matches and preference judgments against ground-truth tutor turns with only small losses in lexical overlap.

What carries the argument

The central object is the steering vector v, a direction in the model's activation space at the final transformer layer, shared across tutors, and the per-tutor coefficient δ_i = exp(u_i) normalized to unit mean. The training signal is a preference objective in the style of bidirectional preference optimization: it favors the actual tutor utterance over the unsteered SFT model's behavior and disfavors the population-mean utterance under steering. The SFT model trained on all tutor utterances supplies the reference 'average tutor' that the vector is defined against.

What would settle it

Regenerate the population-mean baseline with different decoding strategies (greedy vs. nucleus sampling, single vs. multiple candidates) and re-learn the steering vector; if the learned δ_i ordering and the mid-dialogue cosine-similarity gains change substantially, the direction is an artifact of the baseline generation rather than a persona axis.

Watch

Extended reading notes

Core claim

The central claim is that a steering vector v learned by preference optimization, applied as activation steering at the final layer, captures tutor-specific variation across dialogue contexts. For each tutor i a scalar δ_i scales v; training encourages the steered model to increase likelihood of the true tutor utterance and decrease likelihood of a generated population-mean utterance, relative to the unsteered SFT model. At inference, α δ_i v is added to hidden activations. Across 21 tutors, the authors find mid-dialogue cosine similarity of steered output to ground truth rises from 0.385 to 0.426 and an LLM judge prefers steered outputs 58.7% of the time, while ROUGE-L and BLEU stay nearly

Load-bearing premise

The load-bearing premise is that the utterances generated by the supervised fine-tuned model over all tutors form a valid 'average tutor' baseline, so that preferring the real tutor's utterance over this generated baseline isolates that tutor's personal style; the paper does not specify how the baseline utterances are generated.

Editorial extensions

If this is right

Tutor style can be controlled without explicit persona prompts, so a single base model can serve many tutoring styles by adding a scaled vector at inference.
Steering primarily changes phrasing and discourse acts rather than surface lexical overlap: ROUGE-L and BLEU stay roughly flat while semantic metrics improve.
The strength of the persona effect can be dialed at inference via α; increasing α raises preference win rates but lowers lexical similarity, revealing a trade-off.
The learned coefficients δ_i embed interpretable structure: their ordering correlates with consistent differences in tutoring behavior, from scaffolding/rapport to task-completion style.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the central claim holds, the same contrastive construction could be applied to student models, letting simulated tutor–student pairs be matched or mismatched along a learned persona axis.
The δ_i continuum suggests that a single 'tutor-style axis' may serve as a low-dimensional summary of pedagogical style in other dialogue corpora, potentially enabling clustering or regularization of tutor behaviors.
The method's usefulness depends on the population-mean reference being a clean anchor; using a more principled or diverse baseline (e.g., multiple samples or a mixture) would test whether the learned direction is stable.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Here is my read on Lee et al. The paper is a legitimate extension of BiPO: instead of a two-sided steering vector, they learn a single shared direction scaled by per-tutor coefficients δ_i, and apply it to tutor personas extracted from real human tutoring dialogues. That is a real novelty, though a modest one—it borrows all the components (BiPO, persona vectors, DPO) and repurposes them for a specific application. The evaluation does include held-out test dialogues, so the simple circularity concern doesn't land: they are not just re-predicting the training pairs.

What the paper does well: the formulation is clean, the unit-mean constraint on δ_i is a sensible way to remove scale ambiguity, and the qualitative analysis linking low-to-high δ_i to a spectrum from rapport-rich scaffolding to minimal task-completion is genuinely interpretable and plausible.

The soft spots are exactly where the reader put them. First, the population-mean baseline \bar{t} in Eq. (2) is never operationally defined: they don't say whether it's greedy decoding, temperature-1 sampling, beam search, or an average over multiple samples. Since the entire training signal is a contrast between the ground-truth tutor utterance and \bar{t}, an underspecified or degenerate baseline could make the learned direction encode something other than persona. This is not a minor detail; it belongs in Section 3.1. Second, all the headline numbers—CS 0.385 to 0.426 in mid-stage, win rate around 0.587—come from a single generation run with temperature 1.0 and top-p 0.95, and no error bars or significance tests. With differences that small, a few seeds could change the story. The early-stage regression is hand-waved away as greeting behavior; that may be right, but it's asserted rather than shown. The α selection also looks post hoc, chosen to preserve lexical similarity, which weakens the claim that the scaling factor is a principled control.

None of this is fatal. The central idea is sound and the held-out evidence points in the claimed direction. But the paper as submitted isn't ready for a clean accept; it needs a precise definition of the baseline generation, multi-seed evaluation with confidence intervals, and a validation-based α selection.

Who gets value from this: people working on activation steering for educational NLP, and anyone interested in style transfer in LLMs. I'd send it to reviewers, with a clear request for those revisions. If they pin down the baseline, it becomes a solid contribution.

Recommendation: engage with it, but don't take the quantitative claims at face value until the baseline is defined.

Referee Report

3 major / 4 minor

Summary. Using the Question-Anchored Tutoring Dialogues-2k corpus (21 tutors), the paper trains an SFT 'population-mean' tutor model, then forms preference pairs (actual tutor utterance vs. SFT-generated mean utterance). A shared steering vector v and tutor-specific positive coefficients δ_i are optimized with a BiPO-style logistic loss that favors the actual utterance and disfavors the mean utterance under the steered model. At inference, the vector is added to layer-32 activations with scaling αδ_i. Evaluation uses ROUGE-L, BLEU, Sentence-BERT cosine similarity, and Prometheus-Eval win rate against ground-truth tutor utterances. The paper reports improved semantic alignment in middle dialogue turns, a trade-off with lexical similarity, and an interpretable δ_i ordering from rapport-rich scaffolding to minimal task-completion behavior.

Significance. The paper addresses a worthwhile and underexplored problem: controlling tutor-specific style in LLMs without explicit persona prompts. Its proposal—a shared steering direction with per-tutor scalar coefficients learned from human dialogues—is a natural extension of BiPO and activation steering, and the interpretability analysis is a plus. The evaluation uses held-out test dialogues, multiple metrics, and qualitative case studies, which are all strengths. However, the current empirical support is not yet convincing: the contrastive baseline used to define the steering direction is underspecified, the reported gains are modest point estimates with no uncertainty quantification, and the evaluation criterion closely mirrors the training objective. If these issues are resolved, the method could be a useful and practical contribution to personalized tutoring with LLMs.

major comments (3)

[§3.1 / §3.2 (displayed preference objective)] The population-mean utterance \bar{t}_{j,k} is never operationally defined. The text says an SFT model 'is used to generate population-mean utterances' but does not specify the decoding method (greedy vs. sampling), temperature, number of candidates, or any aggregation across stochastic samples. Since the preference objective in §3.2 uses log πθ(\bar{t}_{j,k} | A_L(·)+δ_i v) as the dispreferred term, the learned v and δ_i depend entirely on this baseline. If \bar{t} is degenerate (short, repetitive, or a single random draw), the learned direction will not isolate tutor-specific persona. Please specify the generation protocol and, ideally, show stability of the learned vector and downstream metrics across reasonable baseline-generation choices.
[§5.1, Tables 1 and 2] All reported scores are point estimates from a single stochastic generation run (temperature 1.0, top-p 0.95). The main quantitative successes are modest (mid-stage CS 0.385→0.426; win rate 0.587; overall win-rate range 0.536–0.582), and the early-stage CS drops from 0.392 to 0.321, contradicting the 'across dialogue contexts' claim in the abstract. Without per-tutor variances, confidence intervals, or significance tests, these differences are hard to distinguish from sampling noise. The hand-waved explanation for the early-stage failure ('likely due to greetings') is not sufficient. Please report bootstrap CIs or multiple seeds, and analyze the early-stage failure more rigorously.
[§4.2 / §3.2] The evaluation metrics (CS and win rate) measure alignment with the ground-truth tutor utterance, which is exactly the preferred side of the preference pairs used to train v and δ_i. Held-out test dialogues mitigate direct leakage, but the training objective and the evaluation criterion are still the same kind of target; the LLM judge is asked which output 'more closely aligns with the tutoring style and language used', which is close to the objective being optimized. To demonstrate that the steering vector learns a reusable persona rather than overfitting to the preference signal, add at least one evaluation not defined by proximity to the training target—for example, human judgments against a pedagogical rubric, a held-out set of tutors, or a downstream task measuring tutoring quality—and show the effect is not driven by a few outlier tutors.

minor comments (4)

[§4.3] The paper states that α=0.5 is selected 'for our qualitative analysis' after presenting Table 2. If this choice was made after looking at test-set numbers, the reported results are optimistically biased. Please state the selection protocol explicitly, e.g., using a validation split or a pre-registered value.
[§3.2 / §4.3] The notation πθ(· | A_L(·)+δ_i v) is ambiguous. It should be clarified that this denotes the probability of the full tutor utterance under the transformer with the steering vector added at layer L, and how this probability is computed during training (i.e., whether the added vector is treated as constant or differentiated through).
[Table 3] There is a typo in the Tutor 283 ratio example: 'as qwell' should be 'as well'. Also, the blue highlighting mentioned in the caption is not visible if the paper is printed in grayscale; consider using a symbol or formatting cue.
[Table 4 / §5.2.2] In the order-of-operations example for Tutor 2488, the authors acknowledge that the steered prediction incorrectly assesses the student response as correct. This weakens the blanket statement that 'steering consistently improves alignment'; please temper the claim or provide a more detailed discussion of this failure case.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the held-out test split breaks the identity between the preference training signal and the evaluation metric.

full rationale

The derivation chain is trainable and evaluated on disjoint splits. In §3.2, Eq. (2) optimizes the shared steering vector v and tutor coefficients δ_i so that the steered model assigns higher relative likelihood to the actual tutor utterance than to the SFT 'population-mean' utterance on the validation dialogues. At test (§4.2), the paper measures cosine similarity (CS) and Win Rate of steered generations against ground-truth tutor utterances on a held-out test split: 'We split the dataset into training, validation, and test sets at the dialogue level' and 'For learning v and u_i, we use the validation set.' Because v and δ_i are fit only on validation and the reported numbers are on test dialogues, the improvements are not forced by construction: the preferred side of the training pairs and the evaluation target are the same type of object, but not the same instances. The paper also compares against the unsteered SFT baseline rather than presenting the preference objective as an independent benchmark. The 'population-mean' utterance \bar{t} is not operationally specified in §3.1 (decoding method, temperature, aggregation); this is a real reproducibility/correctness gap, but an underspecified baseline is not circularity. Self-citations [25,26] and the author-affiliated dataset [34] appear in related work and are not load-bearing for the main derivation. The δ_i interpretability analysis is post-hoc qualitative illustration, not an equation-level reduction. No step in the paper's own equations reduces a prediction to a fitted input, so no circular step is identified.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central method rests on a learned vector v, per-tutor scalars δ_i, and a global α chosen on test data. The key domain assumptions are that the SFT model gives a valid population-mean baseline and that the preference pairs encode true tutor-specific style.

free parameters (3)

shared steering vector v = not reported (hidden-dim vector)
Learned via the BiPO-style objective to increase likelihood of target tutor utterances over SFT-generated population-mean utterances; central intervention.
tutor-specific coefficients δ_i (via u_i) = learned u_i mean 0.053, std 0.106; δ_i shown in Figure 2
Scalars controlling per-tutor steering strength; learned on the validation set; their ordering is interpreted as tutoring style.
global scaling factor α = 0.5 selected for qualitative analysis; 0.3, 0.7, 1.0 also reported
Controls strength of steering at inference; appears to be selected after seeing test-set results in Table 2, which is a post hoc choice.

assumptions (3)

domain assumption The SFT model trained on all tutor utterances is a faithful proxy for population-mean tutor behavior.
Invoked in Section 3.1 to construct the dispreferred side of preference pairs; if wrong, the steering direction is not anchored to a meaningful average.
domain assumption For every tutor, dialogue, and turn, the actual tutor utterance is preferred over the SFT-generated population-mean utterance.
This preference is the training signal in Eq. (2); it assumes human tutor utterances are always the desired target, though some examples in Table 3 (e.g., Tutor 2488) are low-quality or confusing.
domain assumption A single shared steering direction v scaled by per-tutor scalar δ_i captures tutor-specific variation.
This representational choice in Section 3.2 assumes persona differences lie on a one-dimensional axis; if they are multidimensional, the method underfits.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Letting Tutor Personas Speak Up for LLMs: Learning Steering Vectors from Dialogue via Preference Optimization." pith.science (2026). https://pith.science/paper/RXYGOS5D

@misc{pith2026260207639,
  author       = {Pith},
  title        = {Pith review of: Letting Tutor Personas Speak Up for LLMs: Learning Steering Vectors from Dialogue via Preference Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RXYGOS5D}},
  note         = {Machine review of arXiv:2602.07639}
}

read the original abstract

With the emergence of large language models (LLMs) as a powerful class of generative artificial intelligence (AI), their use in tutoring has become increasingly prominent. Prior works on LLM-based tutoring typically learn a single tutor policy and do not capture the diversity of tutoring styles. In real-world tutor-student interactions, pedagogical intent is realized through adaptive instructional strategies, with tutors varying the level of scaffolding, instructional directiveness, feedback, and affective support in response to learners' needs. These differences can all impact dialogue dynamics and student engagement. In this paper, we explore how tutor personas embedded in human tutor-student dialogues can be used to guide LLM behavior without relying on explicitly prompted instructions. We train a steering vector using preference optimization: an activation-space direction that guides model responses toward specific tutor personas. We find that this steering vector captures tutor-specific variation across dialogue contexts, improving semantic alignment with ground-truth tutor utterances and increasing preference-based evaluations, while largely preserving lexical similarity. Analysis of the learned scaling coefficients further reveals interpretable structure across tutors, corresponding to consistent differences in tutoring behavior. These results demonstrate that activation steering offers an effective and interpretable way for controlling tutor-specific variation in LLMs using signals derived directly from human dialogue data.

Figures

Figures reproduced from arXiv: 2602.07639 by the authors.

**Figure 2.** Heatmap δi across 21 tutors with different tutoring styles [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 4 linked inside Pith

[1]

Personalized steering of large language models: Versatile steering vectors through bi-directional preference optimization

Yuanpu Cao, Tianrong Zhang, Bochuan Cao, Ziyi Yin, Lu Lin, Fenglong Ma, and Jinghui Chen. Personalized steering of large language models: Versatile steering vectors through bi-directional preference optimization. Advances in Neural Information Processing Systems, 37:49519–49551, 2024

2024
[2]

Livehint overview

Carnegie Learning. Livehint overview. Online: https://support.carnegielearning.com/help-center/ math/livehint/article/livehint-overview/, 2024

2024
[3]

Persona vectors: Monitoring and controlling character traits in language models.arXiv preprint arXiv:2507.21509, 2025

Runjin Chen, Andy Arditi, Henry Sleight, Owain Evans, and Jack Lindsey. Persona vectors: Monitoring and controlling character traits in language models.arXiv preprint arXiv:2507.21509, 2025

arXiv 2025
[4]

From problem-solving to teaching problem-solving: Aligning LLMs with pedagogy using reinforcement learning

David Dinucu-Jianu, Jakub Macina, Nico Daheim, Ido Hakimi, Iryna Gurevych, and Mrinmaya Sachan. From problem-solving to teaching problem-solving: Aligning LLMs with pedagogy using reinforcement learning. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 272–292, November 2025

2025
[5]

Dynamics of affective states during complex learning.Learning and Instruction, 22(2):145–157, 2012

Sidney D’Mello and Art Graesser. Dynamics of affective states during complex learning.Learning and Instruction, 22(2):145–157, 2012

2012
[6]

Gemini.https://deepmind.google/technologies/gemini/, 2023

Google. Gemini.https://deepmind.google/technologies/gemini/, 2023. Accessed: 2026-01-21

2023
[7]

Collaborative dialogue patterns in naturalistic one-to-one tutoring.Applied cognitive psychology, 9(6):495–522, 1995

Arthur C Graesser, Natalie K Person, and Joseph P Magliano. Collaborative dialogue patterns in naturalistic one-to-one tutoring.Applied cognitive psychology, 9(6):495–522, 1995

1995
[8]

LoRA: Low-rank adaptation of large language models

Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. InInternational Conference on Learning Represen- tations, 2022

2022

Show all 35 references

[9]

Supercharge your teaching experience with khanmigo

Khan Academy. Supercharge your teaching experience with khanmigo. Online: https://www.khanmigo.ai/, 2023

2023
[10]

Prometheus 2: An open source language model specialized in evaluating other language models

Seungone Kim, Juyoung Suk, Shayne Longpre, Bill Yuchen Lin, Jamin Shin, Sean Welleck, Graham Neubig, Moontae Lee, Kyungjae Lee, and Minjoon Seo. Prometheus 2: An open source language model specialized in evaluating other language models. InProceedings of the 2024 Conference on...

2024
[11]

Preserving diversity in supervised fine-tuning of large language models

Ziniu Li, Congliang Chen, Tian Xu, Zeyu Qin, Jiancong Xiao, Zhi-Quan Luo, and Ruoyu Sun. Preserving diversity in supervised fine-tuning of large language models. InThe Thirteenth International Conference on Learning Representations, 2025

2025
[12]

Rouge: A package for automatic evaluation of summaries

Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. InText summarization branches out, pages 74–81, 2004

2004
[13]

In-context vectors: making in context learning more effective and controllable through latent space steering

Sheng Liu, Haotian Ye, Lei Xing, and James Zou. In-context vectors: making in context learning more effective and controllable through latent space steering. InProceedings of the 41st International Conference on Machine Learning, ICML’24. JMLR.org, 2024

2024
[14]

Personality-aware student simulation for conver- sational intelligent tutoring systems

Zhengyuan Liu, Stella Xin Yin, Geyu Lin, and Nancy Chen. Personality-aware student simulation for conver- sational intelligent tutoring systems. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 626–642, 2024

2024
[15]

Training millions of personalized dialogue agents.arXiv preprint arXiv:1809.01984, 2018

Pierre-Emmanuel Mazaré, Samuel Humeau, Martin Raison, and Antoine Bordes. Training millions of personalized dialogue agents.arXiv preprint arXiv:1809.01984, 2018

2018 arXiv
[16]

An introduction to the five-factor model and its applications.Journal of personality, 60(2):175–215, 1992

Robert R McCrae and Oliver P John. An introduction to the five-factor model and its applications.Journal of personality, 60(2):175–215, 1992

1992
[17]

Chatgpt.https://chat.openai.com, 2023

OpenAI. Chatgpt.https://chat.openai.com, 2023. Accessed: 2026-01-21

2023
[18]

Autotutor meets large language models: A language model tutor with rich pedagogy and guardrails

Sankalan Pal Chowdhury, Vilém Zouhar, and Mrinmaya Sachan. Autotutor meets large language models: A language model tutor with rich pedagogy and guardrails. InProceedings of the Eleventh ACM Conference on Learning@ Scale, pages 5–15, 2024

2024
[19]

Bleu: a method for automatic evaluation of machine translation

Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. InProceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002

2002
[20]

Towards the pedagogical steering of large language models for tutoring: A case study with modeling productive failure

Romain Puech, Jakub Macina, Julia Chatain, Mrinmaya Sachan, and Manu Kapur. Towards the pedagogical steering of large language models for tutoring: A case study with modeling productive failure. InFindings of the Association for Computational Linguistics: ACL 2025, pages 26291...

2025
[21]

Direct preference optimization: Your language model is secretly a reward model.Advances in neural information processing systems, 36:53728–53741, 2023

Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model.Advances in neural information processing systems, 36:53728–53741, 2023

2023
[22]

Sentence-BERT: Sentence embeddings using Siamese BERT-networks

Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using Siamese BERT-networks. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), p...

2019
[23]

Steering llama 2 via contrastive activation addition

Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Turner. Steering llama 2 via contrastive activation addition. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15504–15522, 2024

2024
[24]

Principles of instruction: Research-based strategies that all teachers should know.American educator, 36(1):12, 2012

Barak Rosenshine. Principles of instruction: Research-based strategies that all teachers should know.American educator, 36(1):12, 2012

2012
[25]

Training llm-based tutors to improve student learning outcomes in dialogues

Alexander Scarlatos, Naiming Liu, Jaewook Lee, Richard Baraniuk, and Andrew Lan. Training llm-based tutors to improve student learning outcomes in dialogues. InInternational Conference on Artificial Intelligence in Education, pages 251–266. Springer, 2025

2025
[26]

Improving the validity of automatically generated feedback via reinforcement learning

Alexander Scarlatos, Digory Smith, Simon Woodhead, and Andrew Lan. Improving the validity of automatically generated feedback via reinforcement learning. InInternational Conference on Artificial Intelligence in Education, pages 280–294. Springer, 2024

2024
[27]

Personality traits in large language models

Gregory Serapio-García, Mustafa Safdari, Clément Crepy, Luning Sun, Stephen Fitz, Marwa Abdulhai, Aleksandra Faust, and Maja Matari´c. Personality traits in large language models. 2023

2023
[28]

Focus on formative feedback.Review of educational research, 78(1):153–189, 2008

Valerie J Shute. Focus on formative feedback.Review of educational research, 78(1):153–189, 2008

2008
[29]

Pedagogical alignment of large language models

Shashank Sonkar, Kangqi Ni, Sapana Chaudhary, and Richard Baraniuk. Pedagogical alignment of large language models. InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 13641–13650, 2024

2024
[30]

The use of worked examples as a substitute for problem solving in learning algebra.Cognition and instruction, 2(1):59–89, 1985

John Sweller and Graham A Cooper. The use of worked examples as a substitute for problem solving in learning algebra.Cognition and instruction, 2(1):59–89, 1985

1985
[31]

Steering language models with activation engineering.arXiv preprint arXiv:2308.10248, 2023

Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J Vazquez, Ulisse Mini, and Monte MacDiarmid. Steering language models with activation engineering.arXiv preprint arXiv:2308.10248, 2023

2023 arXiv
[32]

Trojan activation attack: Red-teaming large language models using steering vectors for safety-alignment

Haoran Wang and Kai Shu. Trojan activation attack: Red-teaming large language models using steering vectors for safety-alignment. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management, CIKM ’24, page 2347–2357. Association for Computing...

2024
[33]

The role of tutoring in problem solving.Journal of child psychology and psychiatry, 17(2):89–100, 1976

David Wood, Jerome S Bruner, and Gail Ross. The role of tutoring in problem solving.Journal of child psychology and psychiatry, 17(2):89–100, 1976

1976
[34]

PIIvot: A lightweight NLP anonymization framework for question-anchored tutoring dialogues

Matthew Zent, Digory Smith, and Simon Woodhead. PIIvot: A lightweight NLP anonymization framework for question-anchored tutoring dialogues. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 27467–27476, November 2025

2025
[35]

Spl: a socratic playground for learning powered by large language model.arXiv preprint arXiv:2406.13919, 2024

Liang Zhang, Jionghao Lin, Ziyi Kuang, Sheng Xu, and Xiangen Hu. Spl: a socratic playground for learning powered by large language model.arXiv preprint arXiv:2406.13919, 2024. 10

2024 arXiv

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

Personalized steering of large language models: Versatile steering vectors through bi-directional preference optimization

Yuanpu Cao, Tianrong Zhang, Bochuan Cao, Ziyi Yin, Lu Lin, Fenglong Ma, and Jinghui Chen. Personalized steering of large language models: Versatile steering vectors through bi-directional preference optimization. Advances in Neural Information Processing Systems, 37:49519–49551, 2024

2024

[2] [2]

Livehint overview

Carnegie Learning. Livehint overview. Online: https://support.carnegielearning.com/help-center/ math/livehint/article/livehint-overview/, 2024

2024

[3] [3]

Persona vectors: Monitoring and controlling character traits in language models.arXiv preprint arXiv:2507.21509, 2025

Runjin Chen, Andy Arditi, Henry Sleight, Owain Evans, and Jack Lindsey. Persona vectors: Monitoring and controlling character traits in language models.arXiv preprint arXiv:2507.21509, 2025

arXiv 2025

[4] [4]

From problem-solving to teaching problem-solving: Aligning LLMs with pedagogy using reinforcement learning

David Dinucu-Jianu, Jakub Macina, Nico Daheim, Ido Hakimi, Iryna Gurevych, and Mrinmaya Sachan. From problem-solving to teaching problem-solving: Aligning LLMs with pedagogy using reinforcement learning. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 272–292, November 2025

2025

[5] [5]

Dynamics of affective states during complex learning.Learning and Instruction, 22(2):145–157, 2012

Sidney D’Mello and Art Graesser. Dynamics of affective states during complex learning.Learning and Instruction, 22(2):145–157, 2012

2012

[6] [6]

Gemini.https://deepmind.google/technologies/gemini/, 2023

Google. Gemini.https://deepmind.google/technologies/gemini/, 2023. Accessed: 2026-01-21

2023

[7] [7]

Collaborative dialogue patterns in naturalistic one-to-one tutoring.Applied cognitive psychology, 9(6):495–522, 1995

Arthur C Graesser, Natalie K Person, and Joseph P Magliano. Collaborative dialogue patterns in naturalistic one-to-one tutoring.Applied cognitive psychology, 9(6):495–522, 1995

1995

[8] [8]

LoRA: Low-rank adaptation of large language models

Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. InInternational Conference on Learning Represen- tations, 2022

2022