Towards Communication-Efficient Social Intelligence in Language Agents

What it claims

TACT trains a social-dialogue agent to reach its goals with fewer tokens by distilling teacher feedback built from tested revisions of the agent's own actions. The paper defines communication cost two ways: action tokens summed over the agent's turns, and total interaction turns by both parties. It argues that shorter wording alone can backfire by triggering clarification turns. On SOTOPIA, Figure 1 puts TACT on the empirical Goal–Cost Pareto frontier, with Goal 5.611 on All and 4.371 on Hard. On mean turns, however, TACT is not the cheapest method.

How it works

1. Rollouts. One student snapshot plays both roles. Only slot A is trained. Both roles refresh with the student between batches, and runs use non-thinking templates.
2. Specialists. For every A action, two frozen specialists each write one alternative. They may share weights but get different instructions. - The expression specialist must keep intent, facts, commitments and action type. - The strategy specialist may change intent, for example by accepting a constraint, proposing another arrangement or ending the interaction.
3. One-step branches. The original action and each candidate get exactly one freshly sampled partner reply from the same partner snapshot and seed. Branches never run to the end of the dialogue.
4. Goal-support proxy. No separate reward model is used. The current student scores the mean log-probability of the raw goal text given the context. IG is the change in that score after the action and the partner's reply.
5. Selection and distillation. The selected revision is shown only to the teacher. The teacher scores the student's original tokens, and the token-level signal is log q_T − log p_old. At deployment the student runs alone.

flowchart TD
  A[Student action a0] --> E[Expression candidate]
  A --> S[Strategy candidate]
  A --> R0[Partner reply r0]
  E --> RE[Reply rE]
  S --> RS[Reply rS]
  RE --> F{Eligible?}
  RS --> F
  R0 --> F
  F -->|none| X[Skip: no update]
  F -->|max IG/tokens| T[Teacher conditioned on reference]
  T --> D[OPD on student's original tokens]

Eligibility rules for a candidate:

• Every candidate needs IG > 0.
• An expression candidate must also be shorter than the original. It can win even if its IG is below the original's.
• A strategy candidate must have IG above the original's, and it may be longer.
• Among eligible candidates, the pick is the one with the highest IG per action token. Ties are broken by a seeded draw.

If the original branch gets no legal reply, that action is skipped.

Evidence

The SOTOPIA figures below come from Figure 1. The panels are labelled All·450 and Hard·70.

• Goal–tokens: TACT's advantage is in Goal per action token, not in turns.
• Turns: Teacher SFT and Vanilla OPD use roughly 3 fewer turns. On Hard, TACT uses more turns than both the initial student and SFT+SDPO.
• AgentSense: the text states only that TACT gains over the initial student while cutting tokens and messages. No numbers appear in the fetched text.
• Ablation: the authors say expression and strategy supervision affect tokens and turns differently. The ablation numbers are not in the fetched text.

Limits

• Truncated page. The fetched text stops partway through the method. Results tables, the base model names and sizes, compute, and any code release are absent.
• No semantic check. Semantic preservation is enforced only by prompt. There is no verified consistency filter, so the authors leave candidates available for manual inspection.
• Proxy, not causation. The goal-support score is explicitly "not a causal attribution" of the dialogue outcome. The eligibility rules are not a non-inferiority test.
• Known failure mode. A concise commitment can raise goal support after a cooperative reply while leaving the agreement ambiguous.
• Noisy signal. Each branch gets a single sampled reply. Seeds make runs reproducible but do not remove response variance.
• Self-play only. Training uses self-play with the student as partner. The fetched text shows no evaluation against human partners.