{"id":"8c4ba9bc-e5c6-4a63-9660-814d683e4c30","arxiv_id":"2505.11958","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"HiPPrO generates counterspeech conditioned on both a strategy and an emotion via hierarchical prefix learning plus preference optimization, and reports gains on a new emotion-labeled corpus.","lead":"The paper introduces HiPPrO, a counter-speech generator that follows both a strategy (informative, denouncing, etc.) and an emotion (joy, anger, etc.) by learning hierarchical prefixes and applying odds-ratio preference optimization. It also releases MultiCONAN, a corpus of 13,973 counterspeech replies annotated with emotion labels.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The ~38% strategy-conformity gain depends on a RoBERTa classifier trained on the same dataset as the generator; without external validation the headline claim is not established.","rationale":"The strongest claim is the ~38% SC improvement; this is what makes the paper notable. That number depends entirely on the RoBERTa SC classifier, so its validity is load-bearing. The reader identified this as the weakest assumption, and I agree. I considered two other candidate concerns: (i) the small human evaluation (30 responses) and (ii) the absence of CounterGEDI, a multi-attribute controllable counterspeech baseline, from Table 2. Both are real limitations: the human study is underpowered and could be cherry-picked, and CounterGEDI is arguably the most relevant prior method. However, neither directly invalidates the headline number the way the SC classifier circularity does. The Rouge gains (≈3%, 2%, 3%) are small and could easily flip with different references, but they are not the paper's main selling point. The SC classifier circularity is concrete: Section 5.2 says the classifier is trained on 'our dataset', i.e., MultiCONAN, and HiPPrO is trained on the same labels; the evaluation therefore measures proximity to a training distribution as much as strategy conformity. The paper's own Limitation section concedes that its metrics don't align with human perception, which is an in-text acknowledgment that the central metric is not trustworthy. The proposed test—independent human annotation of strategy on a larger sample, with source blinding—would settle whether HiPPrO actually produces more strategy-conformant text or just text that mimics the classifier's training signal. Given the human evaluation, though small, does show some preference for HiPPrO over baselines, a conditional acceptance remains appropriate until such validation is done.","tokens_in":25851,"tokens_out":5955,"duration_ms":57882,"concrete_test":"Select a stratified random sample of 300 generated responses (100 from HiPPrO, 100 from GPT-3.5 FS, 100 from GPT-4 FS) on the MultiCONAN test set. Have trained annotators blind to source label each response for strategy category (positive/informative/questioning/denouncing) using the annotation guide from Section 9.2, and compute the proportion matching the intended strategy. If HiPPrO's human-evaluated strategy-conformity advantage over GPT-3.5/4 is not reproduced (e.g., less than the classifier-reported 0.384 gap), the SC-based headline is an artifact of classifier training on the generator's training distribution. Also report Cohen's kappa between the RoBERTa SC classifier and the human labels separately for each model to detect systematic bias against out-of-distribution outputs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim — a ~38% improvement in strategy/intent conformity (SC) — is measured with a RoBERTa-large classifier fine-tuned on MultiCONAN (Section 5.2), the same dataset used to train HiPPrO's prefixes and preference tuning. This creates a circular evaluation: the classifier rewards the exact label distribution and surface style the generator was trained to reproduce, so a high SC score may reflect in-distribution mimicry rather than genuinely more strategy-conformant counterspeech. The risk is amplified by the classifier's 0.86 accuracy and by the large distribution gap between HiPPrO and the GPT baselines: HiPPrO's SC rises by 0.384 while its EC, TC, and CoSim fall (Table 2), consistent with a shift toward MultiCONAN-specific phrasing rather than a general improvement in attribute control. The paper's Limitation section explicitly concedes that its automatic metrics do not fully align with human perception. Until SC is validated by independent annotation or an out-of-distribution classifier, the headline 38% improvement is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces HiPPrO, a two-stage framework for counterspeech generation conditioned jointly on strategy and emotion. In the first phase, the model learns hierarchical prefix vectors on top of FLAN-T5: strategy-specific prefixes first, then emotion-specific prefixes initialized from the strategy prefixes, with the base model frozen. In the second phase, it applies ORPO, a reference- and reward-model-free preference optimization, using the ground-truth counterspeech as the chosen response and the phase-one generation as the rejected response. The authors also release MultiCONAN, an extension of IntentCONANv2 in which all 13,973 counterspeech instances are annotated with one of five emotion labels. Evaluation includes ROUGE, METEOR, BERTScore, cosine similarity, RoBERTa-based strategy/emotion/target conformity scores, toxicity, and a human study; the abstract reports a roughly 38% improvement in strategy conformity and roughly 3%, 2%, and 3% improvements in ROUGE-1, ROUGE-2, and ROUGE-L over baseline models.","tokens_in":26099,"tokens_out":8864,"duration_ms":78448,"significance":"If the empirical claims hold, the paper makes a useful contribution: multi-attribute conditioning through hierarchical prefix learning is parameter-efficient, ORPO avoids the need for a reference model, and the MultiCONAN dataset is a concrete resource for counterspeech research. The paper is commendably transparent in releasing code and data, and it compares against a broad set of baselines including retrieval-based methods, zero-/few-shot LLMs, and prefix-tuning variants, with sensible ablations on virtual-token size and on DPO versus ORPO. However, the central quantitative claim depends on an evaluation classifier trained on the same data as the generator, and the headline improvement figures are reported as absolute differences labeled as percentages. The human evaluation, which the paper itself identifies as important because automatic metrics do not fully align with human perception, is based on only 30 examples. These issues need to be addressed before the headline claims can be accepted.","major_comments":[{"comment":"The headline strategy-conformity gain is measured with RoBERTa-large classifiers fine-tuned on MultiCONAN, the same dataset from which HiPPrO's prefixes and preference pairs are derived. The reported SC of 0.929 for HiPPrO, and the +0.384 improvement over GPT-4 ZS, could therefore reflect in-distribution mimicry of the training labels rather than genuinely better adherence to the intended strategies. The paper's own Limitation paragraph states that the automatic metrics do not fully align with human perception. This concern is load-bearing because the abstract's central quantitative claim (~38% improvement) rests on SC. I request validation with an external classifier or with an independently annotated sample of generated outputs before this claim is accepted.","section":"Section 5.2, Table 2"},{"comment":"The manuscript reports \"~38% improvement in intent conformity\" and \"~3%, ~2%, ~3% improvement\" in ROUGE, but these numbers are absolute differences in metric values, not relative percentages. For example, SC 0.929 minus 0.545 (GPT-4 ZS) equals 0.384, which is an absolute gain of 0.384, not a 38% relative gain; the relative gain against that baseline would be roughly 70%. Similarly, the ROUGE deltas in the last row of Table 2 are absolute point differences. The terminology should be corrected, and the comparison baseline for each delta should be stated explicitly, since the best SC value in Table 2 among all baselines is DialoGPT at 0.714, not the baseline implied by the reported +0.384.","section":"Abstract and Table 2"},{"comment":"The human evaluation is based on only 30 randomly selected responses rated by 35 experts, and the reported win rates (e.g., 0.96 for HiPPrO vs. CoARL) are presented without confidence intervals, inter-rater agreement statistics, or significance tests. Given that the Limitation section acknowledges that automatic metrics do not fully align with human perception, the human evaluation is important evidence. At n=30, the evidence is too thin to support the strong claim that HiPPrO is superior on all four subjective dimensions. I recommend reporting the full rating distribution, agreement statistics, and a confidence interval or significance test for the win rates.","section":"Section 6.3, Table 3"}],"minor_comments":[{"comment":"The terms \"Rouge\" and \"Meteor\" should be \"ROUGE\" and \"METEOR\", and \"odd ratio\" in Equations (3)-(6) and in Figure 2 should be \"odds ratio\".","section":"Throughout"},{"comment":"The HiPPrO row contains \"0.5670.929\" without a separator, which makes the table difficult to read; this appears to be a missing space between CoSim and SC values.","section":"Table 2"},{"comment":"The ablation row labels use \"PrifixTuning\" in several places; this should be \"PrefixTuning\".","section":"Table 2"},{"comment":"Terminology is inconsistent: the abstract says \"intent conformity\" while Section 5.2 and Table 2 use \"strategy conformity\" (SC); please pick one term and use it consistently.","section":"Abstract, Section 5.2"},{"comment":"The computing section says \"NVIDIA RTX A100 with80GB RAM\"; there is a missing space before \"80GB\".","section":"Appendix 9.9"},{"comment":"The reference list contains duplicate entries: Fanton et al. (2021a) and (2021b) are the same paper, and Masud et al. (2021a) and (2021b) are also identical; these should be merged.","section":"References"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a real empirical paper about counterspeech generation, with a new emotion-annotated dataset (MultiCONAN) and a model configuration (HiPPrO) that stacks hierarchical prefix learning on top of ORPO. The work is incremental, as both components come from prior work, but the specific combination and dataset are new, and the authors did a thorough baseline comparison. The headline claim is a ~38% improvement in strategy conformity relative to strong LLM baselines. Treat that number with caution. It comes from a RoBERTa-large classifier fine-tuned on MultiCONAN, the same data used to train the generator. That makes the SC score at least partially self-referential: the model is rewarded for reproducing the exact label distribution it was trained on, not necessarily for producing more strategy-conformant text in a general sense. The stress-test note is right that HiPPrO's SC jumps while its EC and TC drop, which is consistent with drifting toward dataset-specific phrasing rather than a broad improvement in attribute control. What the paper does well: the dataset is a real contribution. The emotion annotations have documented inter-annotator agreement, with Cohen's kappa mostly above 0.7 and some lower pairwise values, and the distribution tables are transparent. The ablations are sensible: dropping ORPO hurts, DPO gives similar results, and the virtual-token sweep shows no gain beyond VT=5. The human evaluation is small, at 30 examples, but the win rates over CoARL and GPT-3.5/4 are consistently high and in the same direction as the automatic metrics. The soft spots are in proportion to the claims. The SC metric is the biggest one; I would want either an independently labeled evaluation set, an out-of-distribution classifier, or a larger human study before trusting the 38% figure. The authors' own limitation section concedes that their automatic metrics do not fully align with human perception. METEOR and CoSim are actually slightly worse for HiPPrO than GPT-3.5 zero-shot, so the improvement is not universal. Table 2 also reports single-point estimates without confidence intervals, so the significance claims rest on per-instance t-tests rather than repeated runs. The paper claims to release code and data; if those artifacts appear, that adds weight. Bottom line: send this to peer review. A serious referee could ask for external SC validation and more human data, but the dataset and method are worth careful review. I would bring it to the reading group and probably cite the dataset.","headline":"A useful new dataset and a reasonable model combo, but the headline strategy-conformity gain is measured by a classifier trained on the same data; send to review, fix the evaluation, and the work stands as a modest contribution.","tokens_in":773,"tokens_out":810,"would_cite":true,"duration_ms":52102,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Multi-attribute conditioning makes counterspeech more strategy-aligned and human-preferred","keywords":["counterspeech generation","hate speech","prefix tuning","hierarchical prefix learning","preference optimization","multi-attribute conditioning","emotion-annotated dataset","controllable text generation"],"falsifier":"Have independent human raters (not the model's authors) label the strategy of HiPPrO outputs without seeing the conditioning token, and compare the match rate to the SC classifier's score; if human-assessed conformity for HiPPrO does not exceed the baselines by roughly 38 percentage points, the reported gain is an artifact of the in-distribution classifier.","tokens_in":25678,"feed_emoji":"🛡️","tokens_out":8535,"duration_ms":73584,"temperature":0.7,"pith_summary":"This paper argues that counterspeech generation becomes more effective when the model is conditioned on two attributes at once — a response strategy such as 'informative' paired with an emotion such as 'joy' — rather than on strategy alone. To test this, it proposes HiPPrO, which learns attribute-specific prefix vectors in a hierarchy (strategy prefixes first, then emotion prefixes on top of them) on a frozen FLAN-T5, and then applies a reward-free and reference-free preference optimization (ORPO) to make outputs less toxic and closer to human-written references. The paper also contributes MultiCONAN, an emotion-annotated extension of IntentCONANv2 covering 13,973 counterspeech instances. It reports a ~38% gain in strategy conformity and small ROUGE gains over several baselines, with human raters preferring HiPPrO outputs over CoARL, GPT-3.5, and GPT-4 few-shot. If correct, this makes automated counterspeech more steerable by human moderators while tuning less than 0.01% of the model's parameters.","feed_headline":"38% better strategy-aligned counterspeech with two-condition prefixes","feed_subtitle":"A two-stage prefix and preference-tuning approach beats LLM few-shot baselines while tuning 0.005% of parameters.","key_machinery":"The central object is the hierarchical prefix stack: two nested sets of tunable key-value prefix vectors, $(\\alpha,\\beta)$ for strategy and $(\\gamma,\\delta)$ for emotion, added to the encoder and decoder of a frozen FLAN-T5. The strategy prefixes are learned first, then the emotion prefixes are trained with the strategy prefixes frozen, so the two attribute conditions compose without retraining the base model. The second piece of machinery is the odds-ratio preference optimization (ORPO) loss, which computes the odds of the ground-truth counterspeech versus the model-generated one and penalizes low-odds outputs; because it works directly on the model's own probabilities, it needs no reward model and no reference model. Together, the hierarchy encodes attribute composition, and ORPO aligns the generated counterspeech with human-written quality while keeping the trainable footprint at 589,824 parameters (0.0052% of the total).","core_discovery":"The paper's central claim is that multi-attribute conditioning, realized through hierarchical prefix learning followed by odds-ratio preference optimization, produces counterspeech that is more strategy-conformant, less toxic, and lexically closer to human references than single-attribute generation models and large instruction-tuned LLMs such as GPT-3.5 and GPT-4. In support, HiPPrO reaches a strategy-conformity score of 0.929 against 0.545 for GPT-4 zero-shot and 0.533 for GPT-3.5 zero-shot, ROUGE-1 of 0.273 against 0.239 and 0.242 for GPT-3.5 zero-shot and few-shot, and wins human evaluations on adequacy, relevance, and argumentative effectiveness against CoARL, GPT-4 few-shot, and GPT-3.5 few-shot (win rates 0.85–0.98). The ablations show that the stacked strategy-plus-emotion prefix space is the driver: flat prefix tuning on FLAN-T5 reaches only 0.470 strategy conformity, removing the ORPO component degrades toxicity and other metrics, and replacing ORPO with DPO slightly hurts performance.","pith_inferences":["The hierarchy encodes an implicit assumption that strategy is the primary attribute and emotion the secondary one; a natural test is whether reversing the order (emotion first, then strategy) preserves the gains, which would clarify whether the stack is compositional or merely additive.","Because the strategy-conformity classifier is trained on the same labels that the model is trained on, the 38% gain may partly reflect in-distribution mimicry; an independent human-labeled strategy set or a classifier trained on a different corpus could measure the true transfer.","If the compositional prefix stack transfers, the same two-stage prefix-adapter approach could be applied to other controllable generation tasks (e.g., style+formality, stance+emotion) where adding a new attribute is as cheap as training a small prefix adapter.","The toxicity scores are small in absolute terms (0.087 for HiPPrO vs 0.030 for GPT-4 zero-shot), so the main measurable benefit of ORPO may be on non-toxicity rather than on the lexical metrics; a dedicated detoxification evaluation on adversarial inputs would clarify this."],"forward_implications":["Multi-attribute conditioning (strategy plus emotion) can be composed through stacked prefixes without retraining the base model, so the same language model can be redirected to new attribute combinations by swapping or extending the prefix stack.","Because only 0.0052% of the model parameters are trained, counterspeech systems can be tailored to specific communities or moderation policies at very low computational cost.","ORPO's freedom from reward and reference models means preference tuning of counterspeech can proceed with just a base model and written references, removing a common scaling bottleneck in aligning generative models.","The strategy-conformity gains suggest that conditioning on intent alone under-specifies counterspeech; emotion is a second, load-bearing axis that affects not only tone but also perceived relevance and adequacy.","MultiCONAN, with 13,973 emotion-labeled counterspeech instances, provides a testbed for studying how emotional tone interacts with strategy in counterspeech generation."],"supporting_citations":[{"why":"Supplies the IntentCONANv2 dataset (13,952 strategy-labeled counterspeech instances) that MultiCONAN extends with emotion labels, and CoARL, the strategy-conditioned state-of-the-art baseline.","marker":"Hengle et al. (2024)"},{"why":"Supplies ORPO, the reward- and reference-free odds-ratio preference optimization used in Phase 2.","marker":"Hong et al. (2024)"},{"why":"Supplies prefix tuning, the parameter-efficient method for adding trainable key-value prefix vectors to a frozen language model.","marker":"Li and Liang (2021)"},{"why":"Supplies the hierarchical prompt learning idea that motivates stacking attribute-specific prefixes.","marker":"Liu et al. (2023b)"},{"why":"Supplies FLAN-T5, the base encoder-decoder model on which HiPPrO's prefixes are trained.","marker":"Chung et al. (2022)"},{"why":"Supplies Generate-Prune-Select (GPS), a three-stage pipeline baseline used for comparison.","marker":"Zhu and Bhat (2021)"},{"why":"Supplies the strategy-conformity evaluation methodology and the intent-conditioned counterspeech generation paradigm.","marker":"Gupta et al. (2023)"},{"why":"Supplies ROUGE, the lexical-similarity metric used for the ROUGE-1/2/L improvement claims.","marker":"Lin (2004)"}],"fun_headline_variants":["38% boost in counterspeech intent conformity with dual-attribute prefixes","Hierarchical prefixes beat GPT-4 for strategy-aligned counterspeech","Two-condition prefixes improve counterspeech by 38% on intent","Multi-attribute prefix learning yields 38% better counterspeech intent","HiPPrO: dual-prefix counterspeech beats GPT-4 with 0.005% tuning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 38% strategy-conformity gain is only meaningful if the RoBERTa-large classifier trained on MultiCONAN truly measures strategy conformity, rather than mostly rewarding outputs that mimic the training distribution.","fun_headline_variants_meta":{"raw":{"variants":["38% boost in counterspeech intent conformity with dual-attribute prefixes","Hierarchical prefixes beat GPT-4 for strategy-aligned counterspeech","Two-condition prefixes improve counterspeech by 38% on intent","Multi-attribute prefix learning yields 38% better counterspeech intent","HiPPrO: dual-prefix counterspeech beats GPT-4 with 0.005% tuning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001021,"raw_usage":{"total_tokens":4354,"prompt_tokens":1039,"completion_tokens":3315,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":655,"completion_tokens_details":{"reasoning_tokens":3216}},"tokens_in":655,"tokens_out":3315,"duration_ms":25163,"temperature":1.0,"reasoning_tokens":3216,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:44:29.472608+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Have independent human raters (not the model's authors) label the strategy of HiPPrO outputs without seeing the conditioning token, and compare the match rate to the SC classifier's score; if human-assessed conformity for HiPPrO does not exceed the baselines by roughly 38 percentage points, the reported gain is an artifact of the in-distribution classifier.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies Generate-Prune-Select (GPS), a three-stage pipeline baseline used for comparison."}],"review_version":1}