A self-driving car merging onto a highway does more than predict the traffic around it. It nudges, and the traffic responds. Most models that predict how other drivers react to a planned maneuver assume, without saying so, that the response is monotone: the harder the ego vehicle presses into a gap, the more likely the other driver yields. Social psychology says otherwise. In assimilation–contrast theory, people move toward a message they find mildly discrepant but push back against one that falls outside their range of acceptance. Opinion-dynamics models formalize this as the "boomerang" or backfire effect. We propose that driver reactions follow the same kind of rule. A modest nudge produces yielding. A nudge past a driver-specific threshold produces defensive acceleration and a refusal to yield. The road layout moves that threshold: a driver with 50 meters of lane left before a lane drop tolerates far less pressure than one with 300.
The note makes one claim. If an action-conditional reaction model is given an explicit, smooth assimilation–contrast response, with a threshold computed from the lane graph, it will predict reactions to aggressive ego plans more accurately than generic learned models. The gain should be largest for ego plans more assertive than those common in the training logs, which is where generic models fall back on the monotone trend they mostly saw. Opinion dynamics supplies the mechanism. The lane-graph environment supplies the variable that moves the threshold. Typed relations between agents (same-lane follower, adjacent-lane, merge conflict) replace proximity heuristics. Multi-agent reaction prediction is the setting where the claim either pays off or fails.
What exists and what is missing
The closest work is Behavior-Inspired Neural Networks for Relational Inference (BINN), which places a nonlinear opinion-dynamics differential equation inside a graph neural network. Each agent carries a latent "preference" that evolves under damping, attention to neighbors, and self-reinforcement, and a decoder turns the preferences into trajectories. BINN supplies our method template. It differs from this proposal in three ways. Its influence term is the standard saturating opinion-dynamics coupling, so it has no contrast or backfire branch. It runs on open courts and particle systems with no map. Its communication structure is a proximity heuristic, and it assumes a fixed number of agents.
On the map side, LaneGCN encodes lanes as a graph with typed predecessor, successor, left and right edges. MultiPath++ encodes the road as polylines. Both handle agent-to-agent interaction with generic attention or gating over logged history, and neither conditions on an ego plan. We reuse a LaneGCN-style encoder as the module that produces thresholds and edge types.
On the interaction side, three papers frame the problem we target. Prediction by Anticipation formalizes action-conditional prediction and the two failure modes it is meant to avoid: freezing when others are assumed never to yield, and collisions when they are assumed never to react. LookOut plans over diverse sampled futures. Towards a Fatality-Aware Benchmark of Probabilistic Reaction Prediction defines reaction prediction on NGSIM ramp merges, collapses reactions into yield/pass prototypes, and scores them with a time-to-collision-weighted Brier score $B_c$. None of these says anything structural about the shape of the response as a function of ego pressure.
Two results from social dynamics make that shape the question worth asking. Opinion Dynamics with Confirmation Bias and Modeling the Assimilation-Contrast Effects in Online Product Rating Systems give explicit non-monotone update rules. When Intuition Fails shows that once repulsive influence exists, interventions that look helpful can backfire at the macro level. This contradicts the guarantee in On the Relationship Between Relevance and Conflict in Online Social Link Recommendations, which holds only under linear dynamics. The traffic analogue is that "more assertive ego plans buy more gaps" may be exactly the intuition that fails.
These papers also impose a constraint. The confirmation-bias work and The Social System Identification Problem both report that hard thresholds create discontinuous, non-convex fitting problems. So the kernel must be smooth by construction.
Core Claim and Model
Each surrounding agent $i$ carries a scalar disposition $x_i(t)$, where positive values mean yielding and negative values mean asserting. The candidate ego plan exerts a pressure $m_i(t) \ge 0$ on agent $i$. We define pressure as the inverse time-to-collision between the ego plan and agent $i$ in their shared conflict region, which is the same quantity the fatality-aware benchmark uses for weighting. The disposition evolves by a BINN-style equation:
$$\dot{x}_i = -d_i x_i + u_i \tanh\Big(\alpha x_i + \sum_j a_{ij} x_j + g(m_i;\tau_i)\Big)$$
Here $d_i$ is damping, $u_i$ is attention gain, $\alpha$ is self-reinforcement, and $a_{ij}$ is a coupling weight between agents. The pressure enters through an assimilation–contrast kernel:
$$g(m;\tau) = \beta\, m\,\sigma\!\left(\frac{\tau - m}{s}\right) - \kappa\, m\,\sigma\!\left(\frac{m - \tau}{s}\right)$$
In this kernel, $\sigma$ is the logistic function, $\tau$ is the acceptance threshold, $s$ is a learned softness, and $\beta,\kappa \ge 0$ are the strengths of the assimilative and contrastive branches. Below $\tau$, pressure pushes the agent toward yielding. Above $\tau$, it pushes the agent toward asserting. Setting $\kappa = 0$ gives a monotone model, which is our key ablation. Because the gates are smooth, gradient-based fitting avoids the discontinuities the identification literature warns about.
The map enters through the threshold. We compute $\tau_i = f_\phi(\text{lane-graph features of } i)$, where $f_\phi$ reads the remaining distance to the lane end, the conflict-zone geometry, and the lane count from a LaneGCN encoder. Relational heterogeneity enters through the couplings: $a_{ij}$ is taken from a small learned table indexed by the lane-graph relation between $i$ and $j$ (same-lane leader or follower, adjacent lane, merge conflict, none), multiplied by a continuous learned gate. This replaces BINN+'s proximity prior, and it also replaces the binary edges that SGSGN itself flags as a limitation.
Agents that enter the scene receive an initial $x_i$ from the encoder, and agents that leave are dropped. This is enough to lift the fixed-cardinality restriction for our purposes without making open systems a research contribution of their own. A decoder maps each disposition trajectory, together with the agent's lane-graph context, to a future trajectory and to a yield/pass probability.
Research Program
The work begins with a falsification step on real data, before any modeling. In NGSIM ramp merges, we bin episodes by the peak pressure the merging vehicle exerted on the mainline vehicle. Within each bin, we estimate the empirical rate at which the mainline vehicle yields, stratified by remaining ramp length. If that rate is monotone non-decreasing in pressure in every stratum, within confidence intervals, then the core premise has no support in this data. In that case, the kernel reduces to an inductive bias for extrapolation only, and we would reframe the project around threshold estimation rather than backfire. The bins are confounded, because aggressive mergers may pick different gaps. We therefore also condition on initial gap and relative speed, and report the result as descriptive evidence, not a causal estimate.
Logs cannot show counterfactual reactions. We therefore build a merge simulator over the NGSIM ramp geometry and the Argoverse lane graph. Its simulated drivers are generated from known assimilation–contrast rules with map-dependent thresholds, mixed with monotone drivers at a controlled proportion $\rho$. Every ego plan can be replayed against the same driver, so the true response curve is available. Training uses logs whose ego plans come from a mild-pressure distribution. Testing sweeps pressure past that range.
The model is then trained on NGSIM merges in the benchmark's protocol and on Argoverse. We probe it in three ways. First, we read out the learned response curve $P(\text{yield} \mid m)$ for each agent. Second, we test whether the learned $\tau$ falls as remaining lane length falls, which is a prediction we commit to in advance. Third, we place the model inside a simple sampling-based contingency planner and measure freezing and collision rates, the failure modes defined by Prediction by Anticipation.
Baselines
The naive reference is a non-reactive model: other agents follow constant velocity, or the action-unconditioned forecast, whatever the ego plan. In a planner this produces the freezing-robot and aggressive-invariance failures. Every reactive model must beat it on $B_c$ and on simulated planning outcomes, or reactivity is not paying for itself.
The published reaction baselines are the HMM, mixture density network and inverse reinforcement learning models from the fatality-aware benchmark. On its NGSIM merge set, the paper reports a Brier score of 0.1099 for the mixture density network, the best plain Brier score, and a fatality-aware $B_c$ of 0.2053 for inverse reinforcement learning, the best safety-weighted score. We run their protocol unchanged, with the same prototype discretization and the same splits. Beating them means a lower $B_c$ than 0.2053 that also keeps plain Brier at or below 0.1099. A lower $B_c$ obtained purely by inflating conservatism $C$ does not count, so we report the $G/C/D$ decomposition.
The strongest competitor is an action-conditional transformer in the spirit of Prediction by Anticipation. It uses the same LaneGCN map encoder and the same agent-history encoder as our model, receives the ego plan as tokens, and is matched to our model's parameter count, with the same hyperparameter search budget and training data. It is the strongest competitor because it can represent a non-monotone response in principle. The only difference between it and our model is whether that shape is imposed. Beating it means lower yield-probability error and trajectory error on held-out high-pressure ego plans in simulation, a lower $B_c$ on NGSIM, and no worse in-distribution accuracy.
As a scale check on Argoverse, we report minFDE and miss rate against LaneGCN, which reports minFDE of 1.36 m at $K=6$ and miss rate 0.16 on the test set. The aim here is to show no regression, not to win. A second dynamics baseline is BINN+ from BINN, with LaneGCN map features concatenated into its environmental input and the ego plan added as an input. This tests whether an existing opinion-dynamics network plus naive map conditioning already gives the benefit.
Three ablations isolate the ingredients. Setting $\kappa = 0$ yields a monotone kernel and tests the backfire claim directly. A single global $\tau$ removes map dependence. Uniform $a_{ij}$ removes relational typing. Each ablation runs with identical training budgets.
Why not just…?
Why not just train a large action-conditional transformer and let it learn non-monotone responses from data? This is the real threat, and we do not dismiss it. With enough logged high-pressure merges, a transformer should recover the curve. Our argument is narrower: assertive merges are rare in naturalistic logs, so a flexible model will interpolate the dominant monotone trend into the region where contrast effects occur. The simulator sweep tests exactly this by training on mild pressure and testing on strong pressure. It also includes a control at $\rho$ near zero, where all drivers are monotone. At that setting our model should match the transformer and not beat it. If the transformer matches us out of distribution as well, the structured kernel adds nothing beyond interpretability, and we will report that.
Why not just use BINN and feed it map features? BINN's coupling is saturating and monotone in its inputs, so concatenating map features cannot produce a sign flip in response to pressure unless the encoder learns one implicitly. The BINN+-with-map baseline measures this directly. If it closes most of the gap to our model, the contribution shrinks to the contrast branch, and the $\kappa=0$ ablation will show how much that branch is worth.
Why not just classify yield versus pass with discrete prototypes, as the fatality-aware benchmark does? Prototype classifiers can be scored on $B_c$, and we include them. But they output a probability for each fixed plan. They encode no relation between plans, so they cannot tell a planner that pressing harder lowers the chance of a yield. The planning experiment, in which we count freezing and collisions when each model scores a sweep of candidate merges, is where this difference should show up.
Why not just learn a recurrent latent intent per edge, as EvolveGraph does? EvolveGraph evolves edge types with a GRU and reports roughly 20% and 27% reductions in minADE/minFDE on H3D. It is unconditioned on ego action, however, and its latent state has no pressure-response structure to inspect or test. We treat it as a secondary forecasting baseline on Argoverse rather than as a reaction model.
Decisive Checks
The idea stands or falls on four results. The NGSIM stratified yield rates must show a non-monotone dip in at least the short-ramp strata. In simulation, our model must beat the matched transformer on held-out high-pressure plans when $\rho$ is large and tie it when $\rho$ is near zero. On NGSIM, $B_c$ must fall below 0.2053 without the gain coming from the conservatism term. And the learned threshold must decrease with remaining lane length, a sign we commit to before looking.
If the first check fails, the backfire premise is unsupported for highway merges, and we would retarget the work to denser settings such as unsignalized intersections. If the second check fails, structure is not needed and the honest conclusion is to use the transformer.
References
- Behavior-Inspired Neural Networks for Relational Inference
- Learning Lane Graph Representations for Motion Forecasting
- MultiPath++: Efficient Information Fusion and Trajectory Aggregation for Behavior Prediction
- Prediction by Anticipation: An Action-Conditional Prediction Method based on Interaction Learning
- LookOut: Diverse Multi-Future Prediction and Planning for Self-Driving
- Towards a Fatality-Aware Benchmark of Probabilistic Reaction Prediction in Highly Interactive Driving Scenarios
- Opinion Dynamics with Confirmation Bias.
- Modeling the Assimilation-Contrast Effects in Online Product Rating Systems: Debiasing and Recommendations
- When Intuition Fails: the Complex effects of Assimilative and Repulsive Influence on Opinion polarization
- On the Relationship Between Relevance and Conflict in Online Social Link Recommendations
- The social system identification problem
- Self-Growing Spatial Graph Networks for Pedestrian Trajectory Prediction
- EvolveGraph: Multi-Agent Trajectory Prediction with Dynamic Relational Reasoning