Condescension is relational. The sentence "don't worry, we'll handle everything for you" is fine as a reply to someone who asked for help. Said unprompted to a refugee, it casts them as a passive recipient of charity. Detecting this kind of patronizing and condescending language (PCL: language that implies the speaker's superiority, often through apparent kindness) therefore needs two things: the speaker's words, and what is known about the person addressed and what they asked for. Today's detectors are trained on isolated news paragraphs. The obvious way to supervise conversations among many language-model agents and users is a single central moderator that rereads the whole transcript at every turn. Its cost grows with transcript length and with the number of participants. The literature also suggests that piling more context on a language model makes subtle social judgments worse, not better.
Control engineers faced the same structure when managing power grids and building networks. A central controller that sees everything becomes intractable. What works instead is giving each subsystem a local controller that exchanges a compact, bounded summary with its neighbors. We propose to carry this idea over. Each conversational agent gets a small local monitor. Before speaking, each agent publishes a short typed commitment of the form "I will keep the agency I attribute to you above level a, assuming your own turns stay within the needs you have disclosed." The core claim is that such contract-style summaries, exchanged only between agents who address each other, let local monitors detect condescension nearly as well as a full-transcript central moderator, at a per-turn cost that stays flat as the group grows. We also claim that the "assumption" half of the contract, which encodes whether help was solicited, is what carries the relational information.
What Exists and What Is Missing
The PCL literature supplies the target concept and the local detectors. Don't Patronize Me! built the paragraph-level dataset. It defined seven categories (among them "Unbalanced power relations," "Compassion," "Presupposition," and "Shallow solution") and reported a RoBERTa binary F1 of 70.63%. The SemEval-2022 Task 4 overview reports top systems in the mid-60s F1 on the binary subtask and the high 40s macro-F1 on categories. Both papers treat PCL as a property of a standalone paragraph. That framing is why "Presupposition" and "Shallow solution" are flagged as needing outside knowledge. In a conversation, much of that knowledge is exactly what the addressee has said, so we reuse these detectors but condition them on partner context.
Two papers constrain our design. In Identifying Power Relations in Conversations using Multi-Agent Social Reasoning, feeding generated arguments directly into a language-model prompt dropped accuracy from 49.0% to 44.3%. Only a structured central factor-graph solver recovered performance, reaching 54.3%, while fine-tuned BERT reached 55.0%. The Computational Anatomy of Humility found that few-shot demonstrations hurt GPT-4 on arrogance and humility coding, which reached 0.70 coarse macro-F1 against a human ceiling of 0.83. Together these rule out "just give the moderator more text." They motivate typed, low-dimensional messages instead of free-text summaries.
Wu et al. are the closest prior work. They decompose argument generation across agents but resolve everything in one global solver. Their task is offline two-speaker classification on six-turn snippets. They do not address growth in participants, real-time monitoring, or vulnerable addressees.
Connotation Frames of Power and Agency scores verbs by the agency and power they assign to their arguments. We reuse this as the coordinate system for the published commitments: how much agency speaker i attributes to addressee j.
From control we borrow the communication architecture, not the theorems.
- Compositional synthesis via assume-guarantee contracts has each subsystem guarantee bounds on its outputs, assuming bounds on its neighbors. Centralized synthesis failed above 100 dimensions, while the compositional method handled 10,000 dimensions in 153 seconds.
- Network Control Using Local Information Exchange has agents send neighbors bounded "forecast sets" of their future states rather than full trajectories. It came within 2% of centralized cost, ran 2.5–5x faster, and was 13% cheaper than fully decoupled control. It also proved zero gap on tree-structured networks. - Distributed Synthesis of Local Controllers shows plug-and-play addition of a subsystem using only its neighbors' summaries.
None of these touches language. Their guarantees rely on linear dynamics and convexity that conversation does not have, so we treat their structure as a design hypothesis to be tested empirically, not as a source of proofs.
Core Claim
Each agent $i$ publishes, for each addressee $j$, an interval commitment $[a_{ij}, \bar a_{ij}]$ on the agency and power its next $k$ turns will attribute to $j$, measured in connotation-frame units. Each addressee $j$ publishes a coarse disclosure descriptor $d_j$: whether it requested help, what kind, and any self-identified group membership, revealed only at the granularity it chooses.
A local monitor at agent $i$ sees only three things: $i$'s own turns, the descriptors of its addressees, and the commitments its addressees made toward it. It flags a turn when either condition holds:
- the turn's measured attribution falls outside $i$'s own commitment, or
- a DPM-trained detector, conditioned on $d_j$, scores it as patronizing.
The commitment is renegotiated locally when $d_j$ changes. For example, an agent that discloses a need relaxes the bound on compassionate phrasing.
Two predictions follow. First, local monitoring should match central monitoring to within a small gap. Second, per-turn monitor cost should stay roughly constant in group size, because each monitor reads a fixed-size message from each of its addressees rather than the full transcript.
Research Program
We build four components.
- Agency/power estimator. A turn-level estimator of attributed agency and power, combining the Sap et al. verb lexicon over dependency parses with a regressor fine-tuned on DPM "Unbalanced power relations" spans.
- Conditioned local detector. A RoBERTa detector fine-tuned on DPM, extended with a disclosure-descriptor input. It is trained on DPM paragraphs paired with synthesized prior turns that do or do not solicit the help described, with labels checked by annotators.
- Commitment protocol. Agents state commitments in a fixed schema before speaking, and monitors verify them.
- Central comparison system. A full-transcript moderator built on the same backbone, so that the only difference between local and central is information access.
We also study mitigation. When a monitor flags a turn, the agent regenerates that turn under its commitment. We measure whether condescension falls without agents becoming curt or withholding help, which is the language analogue of the conservatism that bounded decoupling introduces in control.
Baselines
The naive reference is no monitoring. It gives the base PCL rate of the agent population, which matters because PCL is only about 10% of natural text in DPM.
The decoupled baseline is the DPM-fine-tuned RoBERTa applied to each turn independently, with no messages. It is the analogue of the non-cooperative system in Darivianakis et al. The gap between it and our method measures what communication buys. For context only: on paragraphs, this detector type scored 70.63% binary F1 in Pérez-Almendros et al. For our multi-party benchmark, no reported result was found on Lacuna, so all numbers there will be new.
The strongest competitor is the central full-transcript moderator. It sees strictly more information than any local monitor, so on accuracy it is an upper bound we must approach, not beat. We run it in two forms:
- the same fine-tuned backbone over the full transcript, truncated only by context limit;
- a zero-shot large instruction-tuned model with the DPM category definitions.
On power relations, the closest published structured method is Wu et al.'s central factor graph, at 54.3% on ICSI snippets against 55.0% for BERT and 49.0% for zero-shot prompting. We reimplement it on our pairwise power labels.
We also run a sliding-window central moderator that reads only the last $w$ turns. This is the cheap fix a practitioner would actually deploy.
All systems share the same backbone where applicable, the same training data, and the same hyperparameter search budget of 20 configurations each, tuned on a shared validation split. Cost is reported as tokens read by monitors per turn.
Beating the central moderator means staying within 2 F1 points of it, mirroring the 2% cost gap that was the acceptable bar in network control. At the same time, per-turn monitor cost must grow sublinearly in group size. Beating the sliding window means a statistically significant F1 advantage at matched token cost.
Three ablations isolate the key ingredient:
- removing the disclosure descriptor, which tests whether the assumption side matters;
- removing commitments and keeping only the detector plus descriptor, which tests whether forward-looking bounds matter;
- replacing typed messages with free-text summaries of equal token budget, which tests the "more is less" concern directly.
Why not just…?
Why not just use one central moderator that reads everything? On accuracy it should win or tie, and we expect it to, since it sees a superset of our information. The question is cost and robustness as groups grow. Every turn requires rereading a transcript that grows with length and participant count. The evidence from Wu et al. and Guo et al. suggests that more context does not reliably improve subtle social judgments. The scaling curves from 3 to 48 agents will show whether the central moderator's accuracy stays flat while its cost climbs. If its accuracy and cost both stay acceptable at realistic group sizes, the proposal has no practical point, and we will say so.
Why not just use a sliding window over recent turns? This is the real threat. Most condescension may be decidable from the last few turns, in which case a window gets the benefits of locality without any protocol. Our benchmark therefore deliberately includes long-range cases, in which an addressee discloses a need or group membership many turns before the patronizing reply. It also includes short-range cases, and we report the two strata separately. If the window matches our method on both, the contract machinery is unnecessary and the honest conclusion is that relational context in these conversations is short-range.
Why not just classify each turn with the best paragraph-level PCL detector? That is the decoupled baseline. It cannot tell solicited help from unsolicited saviourism. The "Presupposition" and "Shallow solution" failures noted in the DPM paper suggest this is exactly where it breaks. The descriptor ablation measures the gap on the subset of turns whose label flips depending on whether help was requested.
Why not just run standard message passing on a factor graph, which is already distributed? Loopy belief propagation over Wu et al.'s graph would decentralize computation. But it still requires every agent's arguments and rebuttals as evidence, and it offers no forward-looking bound that an agent can check before it speaks. We include it as an additional competitor on the power-relation task. If it matches our accuracy at similar cost, our contribution narrows to the mitigation side.
Proposed Experiments
The main benchmark is newly constructed. We simulate conversations of $N \in \{3, 6, 12, 24, 48\}$ open-weight model agents, some playing members of the five DPM vulnerable groups, using DPM paragraphs as seed material. Two thousand conversations are annotated turn by turn with the DPM protocol (0–4 scale plus categories), with three annotators per turn and agreement reported. Disclosure timing is varied by design so that long-range dependence can be measured.
Two real-data checks guard against simulation artifacts.
- ICSI meetings. Full multi-speaker ICSI meetings, scored on pairwise power relations, extending Wu et al.'s six-turn snippets to whole meetings.
- Reddit reply threads. The humility/arrogance Reddit threads from Guo et al., scored on arrogance detection. Reply threads are trees, and Darivianakis et al. proved local exchange has zero gap on trees in their setting. We therefore test the analogous prediction that the local-to-central gap is smallest on these threads and largest on the dense, many-to-many simulated groups.
Metrics on all three settings:
- turn-level PCL F1, and macro-F1 over categories;
- the gap to the central moderator;
- monitor tokens and latency per turn as functions of $N$;
- for plug-and-play, F1 on conversations where an agent joins mid-stream, with no retraining or re-prompting of existing agents.
For mitigation, we measure the PCL rate after regeneration and human-rated helpfulness. The goal is fewer condescending turns without loss of substantive help.
Three results are decisive:
- Scaling. At $N=24$ and $N=48$, local monitoring stays within 2 F1 points of the full-transcript moderator while per-turn cost is roughly flat.
- Long-range advantage. On long-range turns, local monitoring clearly beats the sliding window at matched cost.
- Mechanism. Removing the disclosure descriptor collapses performance on the solicited-versus-unsolicited subset, while free-text summaries of equal size underperform typed commitments.
If the first result fails, decomposition is not viable for this problem. If the second fails, the cheap window wins. If the third fails, the claimed mechanism is wrong even if the numbers look good.
References
- Don't Patronize Me! An Annotated Dataset with Patronizing and Condescending Language towards Vulnerable Communities
- SemEval-2022 Task 4: Patronizing and Condescending Language Detection
- Identifying Power Relations in Conversations using Multi-Agent Social Reasoning
- The Computational Anatomy of Humility: Modeling Intellectual Humility in Online Public Discourse
- Connotation Frames of Power and Agency in Modern Films
- Compositional synthesis for linear systems via convex optimization of assume-guarantee contracts
- A Robust Optimization Approach to Network Control Using Local Information Exchange
- Distributed Synthesis of Local Controllers for Networked Systems With Arbitrary Interconnection Topologies