Deep in an online argument, people stop naming what they are arguing about. A reply like "That's not true" or "Exactly this" carries a clear position on Tesla or Bitcoin, but only through the chain of messages above it. Current conversational stance detectors, which classify each post as for, against, or neutral toward a target, read the whole thread at once and learn everything jointly for one topic. They still score only about 50% on deep Reddit threads. We propose to factor the problem into two parts. The first is a reply relation, meaning whether a message agrees with, disagrees with, or ignores the message it answers. The second is an anchor, meaning the nearest earlier message whose stance toward the target is stated outright. A reply's stance is then roughly the anchor's stance with its sign flipped once for every disagreement along the path. The key bet is that reply relations look the same across topics ("you're wrong" means the same thing about Biden as about SpaceX), while anchors are tied to their topic.
That split tells us where to apply adversarial domain adaptation, a training trick that makes features uninformative about which domain they came from. We apply it only to the reply-relation encoder and leave the topic-specific anchor encoder free. The split also turns a subjective task into one with a partly checkable intermediate structure. When several messages in a thread carry gold labels, the implied sign flips between them can be checked. This is the kind of verifiable intermediate step that reasoning-transfer work relies on. The question we test is whether composing topic-invariant relations transfers to unseen topics better than whole-thread encoders, whole-representation adversarial alignment, and large reasoning models prompted with the raw thread.
Core Claim
Cross-topic failure in conversational stance comes from entangling two signals. One is target-agnostic and highly transferable (how a reply relates to its parent). The other is target-bound (what an explicit claim says about the topic). Domain-adversarial training applied to the whole representation hurts transfer because it strips out the target-bound part. Applied only to the relation channel, inside an explicit path composition, it should help. Concretely, for an utterance $u$ with nearest explicit anchor ancestor $a$, the composed prediction is
$$\hat{s}(u) = s(a)\cdot\prod_{e\in \text{path}(a,u)} r(e),$$
where $s(\cdot)\in\{+1,-1,0\}$ is stance toward the target and $r(e)\in\{+1,-1,0\}$ is the relation on reply edge $e$ (agree, disagree, unrelated). A learned gate lets $u$ override the composition when it states its own stance. The training objective is
$$\mathcal{L} = \mathcal{L}_{\text{stance}} + \alpha\,\mathcal{L}_{\text{consist}} - \lambda\,\mathcal{L}_{\text{topic}}(D(h_r)).$$
Here $\mathcal{L}_{\text{consist}}$ penalizes predicted edge relations that contradict the sign flip implied by gold labels on both ends of a path. $\mathcal{L}_{\text{topic}}$ is the loss of a topic discriminator $D$ that sees only the relation encoding $h_r$, trained through a gradient reversal layer. $\alpha$ and $\lambda$ are weights.
This differs from every adversarial stance model we found. Adversarial Domain Adaptation for Stance Detection aligns whole claim–document representations from FEVER to FNC. Adversarial Learning for Zero-Shot Stance Detection on Social Media (TOAD) strips topic tokens from isolated tweets. Cross-Domain Label-Adaptive Stance Detection (MoLE) combines domain experts with gradient reversal across 16 datasets of isolated posts. None of the three models reply structure, and all three make the whole representation invariant. Our closest methodological precedent is Adversarial Topic-Aware Prompt-Tuning for Cross-Topic Automated Essay Scoring (ATOP). It found that full topic invariance destroys topic-bound judgments and split its representation into shared and topic-specific prompts. We borrow that split, but we draw the boundary at a structural unit (the reply edge versus the anchor) rather than at learned prompts.
Research Program
The first step is a cheap falsification study. On a few hundred MT-CSD parent–child pairs, we will hand-label the reply relation and measure how often sign composition from a labeled ancestor reproduces the gold child stance. This tells us the ceiling of the composition prior before anything is built. If composition holds only for a small minority of edges, the gate will do all the work and the idea reduces to a weak auxiliary loss.
The second step is the model. A shared multilingual encoder (XLM-R-large, so that Weibo is reachable later) feeds three heads. The relation head reads a (parent, child) pair with target mentions masked. The anchor head reads a single utterance plus the target and outputs a stance and an "explicitness" score that selects the anchor. The gate decides between the composed and the direct prediction. Anchor selection is soft: we use the explicitness-weighted mixture over ancestors, following the ancestral-lineage view of Improved Target-specific Stance Detection on Social Media Platforms by Delving into Conversation Threads and the ancestor-only attention of Who Did They Respond to? Conversation Structure Modeling Using Masked Hierarchical Transformer. The adversarial discriminator is attached only to the relation head.
The third step is the reasoning-transfer lens, used as a test rather than as decoration. CPL: Critical Plan Step Learning Boosts LLM Generalization in Reasoning Tasks reports that separating abstract plan steps from domain content is what lets math-trained skills reach code and science. Satori: Reinforcement Learning with Chain-of-Action-Thought Enhances LLM Reasoning via Autoregressive Search reports that internalized reflection learned on math transfers to logic and code. Both depend on verifiable rewards, and the file names the lack of verifiers in soft domains as the bottleneck. Our consistency term supplies a partial verifier. We will therefore fine-tune a 7B model with reinforcement learning. Its reward is final-label correctness plus agreement of its written anchor-and-relation plan with labeled ancestors. We run it twice: once initialized from Satori-Qwen-7B and once from its non-reasoning base. This separates three things: whether the decomposition helps a generative model, whether a verifiable intermediate reward helps, and whether math-trained reflection adds anything on top.
Baselines
The naive reference is a fine-tuned encoder on the flattened target utterance plus its ancestors (Branch-BERT style), trained on source topics and tested on the held-out topic. As an in-domain anchor, standard BERT is reported at 44.91% $F_{avg}$ on MT-CSD in A Challenge Dataset and Effective Models for Conversational Stance Detection.
The strongest structural competitor is GLAN from the same paper. It combines global attention, local convolutions, and a graph network over the reply tree, and it reports 50.47% $F_{avg}$ on MT-CSD, ahead of GPT-4 (48.17%) and Llama 2-70b (42.97%). Those are in-domain figures. We found no reported result on Lacuna for GLAN, or any model, under leave-one-topic-out transfer on MT-CSD, so we will re-run GLAN ourselves. For the zero-shot conversational setting, the strongest method is SITPCL from Zero-Shot Conversational Stance Detection: Dataset and Approaches, which reports 43.81% F1-macro on ZS-CSD mixed targets. Our method has to beat that number in ZS-CSD's own protocol.
The strongest adversarial competitor is GLAN or Branch-BERT with a DANN-style gradient-reversal discriminator on the full pooled representation. This is the direct transplant of MoLE or the FEVER-to-FNC approach into threads. The strongest LLM competitors are Satori-Qwen-7B and Qwen-2.5-7B-Instruct, each prompted with the full ancestral thread and chain-of-thought with self-consistency voting, in the style of Stance Reasoner: Zero-Shot Stance Detection on Social Media with Explicit Reasoning. That paper reports 72.6 F1 on isolated SemEval-2016 tweets but did not test threads.
Three ablations isolate the key ingredient. The first removes the adversary entirely, giving composition alone. The second moves the adversary to the full representation, testing whether invariance belongs in the relation channel. The third keeps the relation auxiliary loss but drops the composition, so that relations are only extra features.
For fairness, every system uses the same backbone size within its tier and the same leave-one-topic-out splits. Each tier gets the same hyperparameter search budget, which we fix at 30 trials tuned on a held-out source topic, never on the target. Adversarial systems get identical $\lambda$ schedules. LLM baselines get the same number of sampled chains as the RL variant uses at test time. Beating a baseline means a macro-F1 gain on held-out topics that survives five seeds with non-overlapping bootstrap intervals. It also has to show up on deep turns (depth above 4), where targets are implicit, rather than only on shallow ones.
Why not just…?
Why not just add a domain discriminator to GLAN? This is the obvious combination and the most natural reviewer suggestion. ATOP's experience is direct evidence that whole-representation invariance removes topic-bound information. The file also records over-alignment failures on topic-specific classes. In threads, the topic-bound signal is exactly what lets an implicit reply be tied back to the target. Our second ablation runs this alternative directly. If full-representation adversarial GLAN matches our partial version, the structural split is not needed and we will report that.
Why not just prompt a reasoning model that already learned reflection on math? This is a real threat. Satori and CPL show that math-trained reasoning transfers to logic and code, so the composition might emerge for free. Against it, GPT-4 trails GLAN on MT-CSD and GPT-4o-mini trails SITPCL on ZS-CSD, which suggests unstructured reasoning over long threads is not enough. We settle the question empirically by comparing prompted Satori with prompted base Qwen, and both with our RL-with-consistency variant. If prompted Satori already closes the gap on deep turns, the right conclusion is that math RL transfers to discourse, and our encoder model is unnecessary.
Why not just fine-tune on more conversational data? The USDC paper (USDC: A Dataset of User Stance and Dogmatism in Long Conversations) says models trained on its Reddit threads transfer well to MT-CSD, so data alone might suffice. We will add USDC as extra source data to both the naive baseline and our model. If the gap disappears under the larger data budget, the decomposition is buying sample efficiency rather than transfer.
Proposed Experiments
The main experiment is leave-one-topic-out transfer over MT-CSD's five topics (Tesla, SpaceX, Trump, Biden, Bitcoin), reporting macro-F1 broken down by turn depth. The secondary experiments are ZS-CSD in its native zero-shot protocol, which includes the claim-versus-noun-phrase split where SITPCL degrades, and Reddit-to-Weibo transfer as a stress test. Tension A in the file asks whether social media resists transfer intrinsically, as MoLE's isolated clusters suggest, or only when posts are treated as isolated sentences. We address it by plotting transfer gain against thread depth. If the composition model's advantage grows with depth, that supports the structural explanation.
Four checks would make or break the claim. First, the hand-labeled composition-validity rate from the falsification study must be high enough to matter. Second, a probe on the relation encoder should recover held-out topic identity at near chance under our model but not under the no-adversary ablation, while anchor-head accuracy stays unchanged. Third, partial adversarial alignment must beat full-representation alignment on held-out topics, especially on deep, target-implicit turns. Fourth, the RL variant with the consistency reward should beat label-only RL from the same initialization; if it does not, the verifiable-intermediate argument fails. If composition validity is low, or full alignment wins, we will report that topic-invariant reply relations are not the right unit, which is itself a useful negative result for adversarial stance adaptation.
References
- Adversarial Domain Adaptation for Stance Detection
- Adversarial Learning for Zero-Shot Stance Detection on Social Media
- Cross-Domain Label-Adaptive Stance Detection
- Adversarial Topic-Aware Prompt-Tuning for Cross-Topic Automated Essay Scoring
- Improved Target-Specific Stance Detection on Social Media Platforms by Delving Into Conversation Threads
- Who Did They Respond to? Conversation Structure Modeling Using Masked Hierarchical Transformer
- CPL: Critical Plan Step Learning Boosts LLM Generalization in Reasoning Tasks
- Satori: Reinforcement Learning with Chain-of-Action-Thought Enhances LLM Reasoning via Autoregressive Search
- A Challenge Dataset and Effective Models for Conversational Stance Detection
- Zero-Shot Conversational Stance Detection: Dataset and Approaches
- Stance Reasoner: Zero-Shot Stance Detection on Social Media with Explicit Reasoning
- USDC: A Dataset of \underlineUser \underlineStance and \underlineDogmatism in Long \underlineConversations