Stress-Gated Recruitment of Specialist Agents for Dialogue Policies That Get Stuck

Conversational agents that pursue a goal, such as comforting a distressed user, closing a negotiation, or persuading someone to donate, tend to fail in one characteristic way. Once a conversation leaves the expected path, they keep repeating the same safe move: another open question, another vague sympathy line, another restatement of the offer. Adding a team of specialist language-model agents that debate each turn helps with reasoning in other domains, but running that team on every turn is expensive. Most turns do not need it, and coordination costs can erase the benefit. This note proposes a middle path. A cheap single-policy agent handles ordinary turns. A slowly accumulating "stuck" signal decides when to bring in specialist agents who propose different moves, and a quality-control rule later removes the specialists that are not earning their cost.

The design borrows one specific control pattern from cell biology. Cells build new mitochondria (their energy-producing compartments) only after a stress signal has persisted, not in response to a single blip. They also break down damaged mitochondria, so capacity tracks sustained demand. We use this as an engineering hypothesis, not as a claim about biology. The hypothesis is that deliberation capacity in a dialogue agent should be added when a stress signal integrated over several turns crosses a threshold that depends on how much budget is left, and removed when it stops paying off. The payoff to test is fewer rigidity failures than a single-policy agent at a small fraction of the cost of an always-on multi-agent team. We also test whether the "integrate over turns, with hysteresis" part matters, or whether a plain per-turn router does just as well.

Core Claim

The most competitive recent approach, DialogXpert, has a frozen language model propose a small set of candidate actions each turn and a small value network pick one. It reports over 94% success with a Vicuna 13B prior and above 97% with Qwen 2.5 14B across negotiation, emotional support and tutoring benchmarks. Its conversations average under 3 turns. Its weak point is structural: the value network can only choose among the candidates the single frozen model proposes. Chain of Strategy Optimization documents that such models carry a strong preference bias toward a few generic strategies. When the right move is outside that biased candidate set, a better value function cannot recover it.

Our claim is that the missing ingredient is conditional diversification of the candidate set. Specialist agents should widen the options only when the conversation shows sustained signs of being stuck, and a budget-aware threshold should keep coordination overhead below the level at which ORCHESTRATE finds decomposition stops paying for itself (about 15% of the token budget). This differs from prior work in three ways:

  • Chain of Strategy Optimization uses a multi-agent team only offline, to generate preference data, and runs a single model at inference.
  • The Adaptive Memory and Reflection system triages single-turn medical questions by complexity, one query at a time, with no notion of stress accumulating across a conversation.
  • The bio-autonomic multi-agent framework, the closest biologically inspired system, uses immune-style danger signals to detect faulty agents on code and math tasks. It does not allocate deliberation capacity by sustained demand, and it does not handle dialogue.

Research Program

The base agent follows the DialogXpert recipe. A frozen language model proposes $k$ candidate dialogue actions, and a value network selects one. On each turn $t$ we compute a raw rigidity signal $r_t \in [0,1]$ from three observable quantities:

  • the semantic similarity between the agent's chosen strategy and its previous choices (self-repetition);
  • the change in the value network's estimate of the chosen action (stagnation);
  • the output of a lightweight classifier for user resistance or worsening emotion, reusing DialogXpert's emotion tracker.

The stress state integrates this signal:

$$ s_t = \lambda s_{t-1} + (1-\lambda) r_t, $$

Here $\lambda \in [0,1)$ sets how many turns of evidence the agent remembers. Specialist recruitment switches on when $s_t > \theta_{\text{on}}(b_t)$ and off when $s_t < \theta_{\text{off}}$, with $\theta_{\text{off}} < \theta_{\text{on}}$. The variable $b_t$ is the fraction of the conversation's token budget already spent, and $\theta_{\text{on}}$ rises linearly in $b_t$, so recruitment becomes harder as money runs out.

The gap between the two thresholds is the hysteresis that prevents flapping between modes. It plays the role of the biological requirement for persistent stress before committing to new capacity. We follow the sirtuin-controlled biogenesis and quality-control coupling described in the dexmedetomidine hepatic ischemia study and the review of SIRT1 in stroke. We borrow only the control motif.

When recruitment is on, three specialists run in parallel, each on the same backbone model but with a different role prompt:

  • A user-state analyst writes a short hypothesis about what the user actually needs.
  • A counter-strategist is instructed to propose actions outside the strategies used in the last few turns, which targets preference bias directly.
  • A critic flags which existing candidates repeat failed moves, in the generator-critic style of MACM.

Their proposals are added to the candidate pool. The same value network then chooses, so the decision rule is unchanged and only the options grow.

The quality-control rule tracks, for each specialist, how often its proposals are selected and whether selected proposals raise the value estimate on the next turn. A specialist whose running utility falls below a floor is dropped before the stress state itself decays. This mirrors the removal of damaged mitochondria and prevents paying for a role that is not helping in the current conversation.

The parameters $\lambda$, $\theta_{\text{on}}$, $\theta_{\text{off}}$, the budget slope and the pruning floor are tuned on a development split with a fixed search budget. The value network is retrained with specialists present, so it learns to evaluate their proposals.

Baselines

The naive reference point is the frozen backbone prompted to act as a goal-directed agent with a strategy list and chain-of-thought, with no value network and no specialists. This measures how much of any gain is due to structure at all.

The strongest existing method is DialogXpert, run with the same backbones it reports (Vicuna 13B and Qwen 2.5 14B), the same user simulator, and the same value-network training budget. Its reported figures are over 94% and above 97% success with under 3 turns on average. A second strong reference is a Chain of Strategy Optimization model. That paper reports strategy macro-F1 on ESC-Pro rising from about 30% under supervised fine-tuning to 34% for LLaMA-3.1-8B. We use it both as a standalone agent and as the proposal model inside our system, to test whether preference-trained proposers make recruitment unnecessary.

An always-on multi-agent baseline runs all three specialists on every turn. It shows the ceiling of the approach and its full cost. The key ablations isolate the proposal's ingredients:

  • A per-turn triage router in the style of the Adaptive Memory and Reflection system recruits specialists based on $r_t$ alone, without accumulation or hysteresis.
  • Random recruitment at the same average rate controls for sheer frequency.
  • Matched-budget extra sampling spends the recruited tokens on drawing more candidates from the single frozen model at higher temperature, instead of on role-prompted specialists.
  • Two further variants remove the pruning rule and the budget coupling, respectively.

All systems share backbones, simulators and a fixed total token budget per conversation. Every tuned component gets the same number of development-set trials.

Beating the strongest method means two things. First, higher success, and fewer self-repetition and termination errors, than DialogXpert on the rigidity stress splits described below. Second, matching it within noise on the standard splits, with coordination overhead below 15% of tokens. Beating the ablations means the stress-gated system outperforms the per-turn router and matched-budget sampling at equal cost.

Why not just…?

Why not just route each turn by a confidence threshold, as the medical triage system does? This is the most serious alternative, and our design can collapse into it: setting $\lambda = 0$ and $\theta_{\text{off}} = \theta_{\text{on}}$ recovers it exactly. We expect it to fall short in two ways. It will recruit on isolated noisy turns, where one sharp user reply is not a stuck conversation. It will also drop specialists the moment one turn looks fine, causing mode flapping. That evidence would appear as higher recruitment rates at equal success, or lower success at an equal rate. If the per-turn router matches our system at equal cost, we will report that the biological timing motif adds nothing and keep only conditional diversification as the contribution.

Why not just sample more candidates from the one model? Extra samples from a single biased model may stay inside the same few strategies. Chain of Strategy Optimization reports that preference bias persists even in frontier models such as GPT-4o and Claude 3.5 Sonnet, which its fine-tuned 8B and 9B models outperformed on strategy accuracy. The matched-budget sampling baseline tests this directly. We will measure strategy entropy of the candidate pools alongside success. If temperature sampling matches role-prompted specialists, the specialist roles are unnecessary.

Why not just fix the proposer with preference fine-tuning and skip specialists? Chain of Strategy Optimization lifts strategy F1 only from about 30% to 34%, which leaves most strategy decisions wrong. Fine-tuning also bakes in behaviour for the training distribution rather than reacting to a conversation that has gone off-script. We run our gate on top of a Chain of Strategy Optimization proposer. If recruitment adds nothing there, the honest conclusion is that better training subsumes it.

Why not just use a bigger proposal model? DialogXpert's move from 13B to 14B-class priors already pushes success above 97%, so scale helps on standard splits. Our question is whether scale fixes the stuck case. We compare our system on a 13B backbone against DialogXpert on the larger backbone, at matched token cost.

Proposed Experiments

The benchmarks are those used by DialogXpert: ESConv, CIMA, CraigslistBargain and P4G, with ExTES as an unseen-distribution test. ESC-Pro serves for strategy-selection F1.

Because reported success is already near ceiling and conversations are very short, standard splits cannot show much difference. Under 3 turns is also suspiciously brief for emotional support, and that is one reason we rely on human evaluation. We therefore build rigidity stress splits by giving the user simulator scripted behaviours taken from the failure modes documented for fixed-agenda systems:

  • topic digressions;
  • repeated pushback after the agent's first strategy;
  • interruptions that reverse an earlier statement;
  • constraint revelations, such as the financial hesitation that motivates foot-in-the-door tactics in the donation persuasion system.

The metrics are:

  • success rate and turns to success;
  • a self-repetition rate over chosen strategies;
  • premature and missed termination rates, the critic failures DialogXpert itself reports;
  • strategy entropy;
  • the fraction of tokens spent on coordination.

We add human pairwise judgments on ESConv for comforting and overall helpfulness, since simulator success alone can reward rushing.

The decisive checks are three:

  1. On stress splits, the stress-gated system must beat DialogXpert in success and in self-repetition while spending under 15% of tokens on coordination. If it cannot, the core claim fails.
  2. It must beat both the per-turn router and matched-budget sampling at equal cost. Failing the first comparison removes the case for integration and hysteresis; failing the second removes the case for specialist roles.
  3. On standard splits it must not lose to DialogXpert, which shows that recruitment stays mostly dormant when conversations go well.

We will also plot recruitment rate against turn index and simulator behaviour. We expect recruitment to concentrate after scripted pushback and to be near zero in cooperative conversations.