In-Context Examples Matter: Improving Emotion Recognition in Conversation with Instruction Tuning

Bo Zhang; Hui Ma; Jinpeng Hu; Zenglin Shi

In-Context Examples Matter: Improving Emotion Recognition in Conversation with Instruction Tuning

T0 review · 0 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read InitERC claims that a one-stage, in-context instruction tuning framework—retrieving illustrative conversation examples and teaching the model speaker-context-emotion alignment in a single pass—substantially outperforms prior two-stage instr

desk verdict Plausible but unverifiable from the abstract alone; the one-stage in-context tuning idea for ERC is worth a look if the full experiments hold up. read the letter →

arxiv 2508.11889 v1 pith:6MCULZQX submitted 2025-08-16 cs.CL

Hui Ma , Bo Zhang , Jinpeng Hu , Zenglin Shi This is my paper · ORCID

classification cs.CL

keywords emotionrecognitioninconversationinstructiontuningin-contextlearninglargelanguagemodelsspeaker-context-emotionalignmentdemonstrationselectionpromptengineering

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

The paper tries to establish that emotion recognition in conversation is better approached with a one-stage instruction tuning method that uses in-context examples as demonstrations, rather than the prevailing two-stage pipeline that first teaches the model speaker characteristics and then context-awareness. It proposes InitERC, which selects relevant dialogue examples from a pool, arranges them in a prompt, and tunes the model to align speaker identity, context, and emotion in one pass. The authors report substantial gains over state-of-the-art baselines on three standard datasets. If correct, this shows that a single joint alignment step can outperform the staged approach, and that the choice of in-context examples—retrieval strategy, ordering, and number—systematically matters for ERC performance.

What carries the argument

The central object is the in-context demonstration set: a small number of retrieved (utterance, context, emotion) example triples placed in the prompt before the target utterance. It carries the argument because it is the mechanism that encodes speaker-context-emotion alignment; the tuning objective teaches the model to imitate these examples rather than to learn speaker and context in separate stages. The paper's design variables—retrieval strategy, example ordering, and number of examples—are all manipulations of this demonstration set, and the reported performance differences are attributed to them.

What would settle it

A controlled comparison in which the same InitERC model is evaluated with randomly selected in-context examples instead of retrieved ones, keeping everything else fixed, and the performance gap disappears, would undermine the claim that retrieval-driven alignment matters. Likewise, if a directly comparable two-stage baseline that adopts the same retrieval, ordering, and example count matches or beats InitERC, the one-stage advantage would be called into question.

Watch

Extended reading notes

Core claim

The central claim is that jointly aligning speaker characteristics, conversational context, and emotion labels through in-context instruction tuning is more effective than separating the task into a speaker-characteristic tuning stage followed by context-aware emotion tuning. The method constructs a demonstration pool, retrieves in-context examples for each utterance, designs a prompt that includes those examples, and tunes the large language model to predict the emotion of the target utterance. On the three widely used ERC benchmarks, this one-stage setup produces substantial improvements over prior state-of-the-art methods. A secondary claim is that the composition of the demonstration set

Load-bearing premise

The load-bearing premise is that the specific retrieval strategy, ordering, and number of in-context examples trained on transfer to new speakers and domains, and that the one-stage joint alignment—not some incidental detail of the implementation—is what produces the gains.

Editorial extensions

If this is right

ERC systems can be built with a single instruction-tuning pass, simplifying training pipelines currently split into two stages.
Retrieval strategy for demonstrations becomes a performance-relevant design choice; choosing relevant examples can matter as much as model size or prompt wording.
The systematic study of retrieval strategy, ordering, and example count provides practical recipes for building ERC prompts.
If the alignment claim is right, the same one-stage in-context approach could transfer to other utterance-level conversational tasks where speaker and context interact, such as dialogue act or sentiment detection.
Reporting consistent gains across three different datasets suggests the approach is not tied to a single benchmark's quirks.

Reading between the lines

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

An implicit testable extension: the same retrieval and ordering recipe might work even without instruction tuning, in pure few-shot prompting—if the demonstrations carry most of the signal, the tuning step could be secondary.
The paper leaves open whether a personalized demonstration pool, built from the target speaker's past utterances, would outperform a generic pool; if speaker-emotion alignment is learned from examples, speaker-specific demonstrations could sharpen it.
Ablating retrieval on held-out speakers or unseen domains would clarify whether the learned alignment generalizes or overfits to benchmark-specific regularities.
The method could extend to other label-rich conversational tasks where two-stage tuning is costly; the paper only demonstrates emotion recognition.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Quick take: I can only see the abstract, so everything here is provisional. The paper's core idea—replacing two-stage instruction tuning for ERC with a single in-context stage where speaker, context, and emotion alignment is learned from demonstrations—is a reasonable, testable hypothesis, and the systematic variation of retrieval strategy, ordering, and example count is exactly the kind of ablation I want to see in this area. If the three-dataset improvement is real and fairly matched against the baselines, this would be a useful contribution to the ERC subfield, not a paradigm shift.

Credit where due: the abstract is honest about what it does: it builds a demonstration pool, selects examples, designs prompts, and tunes on those in-context examples. That's not a new mechanism, but applying it specifically to the speaker-context-emotion alignment gap in one stage is a sensible move. The claim that multi-stage tuning constrains joint alignment is plausible and worth testing.

Soft spots: the abstract gives no numbers. 'Substantial improvements' is meaningless without effect sizes, error bars, or at least a table. I also can't tell whether the baselines are matched for compute, parameter count, or number of training examples. The three factors studied (retrieval, ordering, count) could easily be tuned on the same benchmarks they evaluate on; the abstract doesn't say whether the selection was done on a validation split. That's a mild fitting-to-benchmark risk, not an indictment. And because I cannot see the full text, I can't verify that the one-stage setup actually captures the alignment better rather than just being a different prompt format.

On the stress-test note: I agree there's no load-bearing flaw visible from the abstract alone. The reader's 'unverified' verdict is the right one. The weakest assumption is the generalizability of the three factors, but that's an empirical question the paper claims to address.

Bottom line: this is a paper for ERC and instruction-tuning folks. If the full text has the standard experimental detail (tables, CIs, ablation, baseline fairness), it deserves a serious referee. I would not cite it from the abstract alone, but I would put it on the reading group list once we get the full version. Send it to peer review; the question is whether the experiments back the claim.

Referee Report

0 major / 3 minor

Summary. The paper proposes InitERC, a one-stage in-context instruction tuning framework for emotion recognition in conversation (ERC). It claims that existing multi-stage instruction tuning methods constrain the joint modeling of speaker characteristics and conversational context, and that InitERC improves speaker-context-emotion alignment by using in-context examples. The framework comprises demonstration pool construction, in-context example selection, prompt template design, and in-context instruction tuning. The authors conduct a comprehensive study of three factors in in-context examples: retrieval strategy, example ordering, and number of examples. They report substantial improvements over state-of-the-art baselines on three widely used ERC datasets, based on the abstract.

Significance. If the claimed results hold, InitERC would provide a simple and effective alternative to the dominant multi-stage instruction tuning paradigm for ERC. The systematic exploration of retrieval strategy, ordering, and number of in-context examples is a useful empirical contribution that could inform future prompt-design and tuning choices for LLM-based emotion recognition. The one-stage design is conceptually appealing and potentially more parameter-efficient than multi-stage pipelines. However, because the full text is unavailable, the reliability of the empirical claims cannot currently be assessed; the abstract alone does not provide effect sizes, statistical confidence, dataset details, or a comparison protocol.

minor comments (3)

[Abstract] The abstract states 'substantial improvements' without reporting concrete numbers, effect sizes, or confidence intervals. For a claims-driven abstract, at least the names of the three datasets and the main evaluation metric should be included to make the result falsifiable and comparable.
[Abstract] The phrase 'comprehensive study on three key factors' is too vague. It is unclear whether the study reports ablations, sensitivity analyses, or statistical tests. Please specify the experimental design in the abstract or refer to a results table in the full text.
[Abstract] The term 'in-context instruction tuning' is central to the contribution but is not defined within the abstract. A brief clarification (e.g., how it differs from standard instruction tuning) would help readers assess the novelty at a glance.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity detectable from abstract-only review

full rationale

The paper as presented in the abstract is an empirical systems paper: it proposes a framework (InitERC) and reports experimental improvements over baselines on three ERC datasets. There is no derivation chain, no equation, and no parameter fitted to a subset of data that is then presented as a prediction. The abstract makes no self-citations and invokes no prior uniqueness theorem or ansatz as load-bearing. Claims such as 'one-stage in-context instruction tuning adapts LLMs to learn speaker-context-emotion alignment' are architectural/empirical claims, not mathematical reductions. The reader's concern about hyperparameters (retrieval strategy, ordering, number of examples) being tuned on the same benchmarks is a legitimate experimental-validity caveat, but it is not circularity under the given criteria: there is no evidence that any reported result is forced by construction from the inputs. Without the full text, no specific circular step can be quoted or exhibited. Therefore the honest finding is no significant circularity, score 0.

Assumptions & free parameters 3 free parameters · 2 assumptions · 0 invented entities

The abstract mentions three design factors that are likely tuned per dataset, which are the main free parameters. No new models, forces, or formal constructs are introduced. The axioms are the standard domain assumptions for any LLM instruction tuning study.

free parameters (3)

number of in-context examples per prompt = not reported in abstract
The paper studies the number of examples as a key factor, implying this is a hyperparameter adjusted to the datasets.
retrieval strategy = not reported in abstract
The choice of retrieval strategy for selecting demonstrations is one of the three factors studied, so it is a design choice fit to the evaluation data.
example ordering heuristic = not reported in abstract
The order of in-context examples is studied as a factor, indicating it is tuned or selected based on performance.

assumptions (2)

domain assumption Large language models can improve emotion recognition in conversation through in-context instruction tuning.
The abstract assumes this premise as the foundation of the proposed method.
domain assumption The three standard ERC datasets used in the experiments are appropriate and representative benchmarks for the task.
The abstract states experiments on 'three widely used datasets' without further justification, treating them as accepted evaluation grounds.

how reviews work

0 comments

Cite this review

Pith. "Pith review of In-Context Examples Matter: Improving Emotion Recognition in Conversation with Instruction Tuning." pith.science (2026). https://pith.science/paper/6MCULZQX

@misc{pith2026250811889,
  author       = {Pith},
  title        = {Pith review of: In-Context Examples Matter: Improving Emotion Recognition in Conversation with Instruction Tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6MCULZQX}},
  note         = {Machine review of arXiv:2508.11889}
}

read the original abstract

Emotion recognition in conversation (ERC) aims to identify the emotion of each utterance in a conversation, playing a vital role in empathetic artificial intelligence. With the growing of large language models (LLMs), instruction tuning has emerged as a critical paradigm for ERC. Existing studies mainly focus on multi-stage instruction tuning, which first endows LLMs with speaker characteristics, and then conducts context-aware instruction tuning to comprehend emotional states. However, these methods inherently constrains the capacity to jointly capture the dynamic interaction between speaker characteristics and conversational context, resulting in weak alignment among speaker identity, contextual cues, and emotion states within a unified framework. In this paper, we propose InitERC, a simple yet effective one-stage in-context instruction tuning framework for ERC. InitERC adapts LLMs to learn speaker-context-emotion alignment from context examples via in-context instruction tuning. Specifically, InitERC comprises four components, i.e., demonstration pool construction, in-context example selection, prompt template design, and in-context instruction tuning. To explore the impact of in-context examples, we conduct a comprehensive study on three key factors: retrieval strategy, example ordering, and the number of examples. Extensive experiments on three widely used datasets demonstrate that our proposed InitERC achieves substantial improvements over the state-of-the-art baselines.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Empathic Prompting: Non-Verbal Context Integration for Multimodal LLM Conversations
cs.HC 2025-10 conditional novelty 4.0 of 10

A multimodal chatbot framework that injects real-time facial-expression-derived valence, arousal, and emotion labels into LLM prompts can condition responses on non-verbal affect.

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.