When people negotiate, they rarely behave as game theory predicts. Backward induction solves a multi-round game by starting at the deadline and reasoning back to the first move. On the formal models, it says a well-informed pair should settle immediately on an efficient package. Real people chat, justify, stall, and sometimes walk away from deals that were worth taking. Mixed-methods research is the standard tool for explaining gaps like this. It pairs numbers, such as outcomes and surveys, with close reading of what people actually said. Its bottleneck is attention: a qualitative analyst cannot carefully code every turn of thousands of dialogues.
This note proposes to let backward induction decide where the qualitative analyst looks. Some negotiation corpora record both players' private values. For those, an analyst can compute, turn by turn, how much value each move gave up relative to the best continuation from that point. The turns with the largest losses are the moments where a person departed most from rational play. Qualitative coding concentrates there, and the codes that come out are fed back into the game as an estimated "relational" payoff term. The test is whether this loop explains held-out human behavior better than game theory alone, qualitative coding alone, or a black-box predictor. Its outputs are an explanatory codebook and a revised game model.
Core Claim
Backward induction can serve as a sampling instrument for qualitative analysis, rather than as a model of what people do. The per-turn value loss it defines picks out moments that are informative about human strategic reasoning. Under a fixed coding budget, codes induced from those turns should predict breakdowns and post-negotiation satisfaction better than codes from randomly chosen turns, from all turns, or from turns flagged by an outcome predictor.
Once those codes enter the terminal payoff, the resulting model should predict human accept/reject decisions and final splits better than the plain equilibrium. This addresses two tensions at once. The first is that the bargaining theory in the file treats delay as pure waste. The second is that coded human negotiation data suggest pre-settlement talk does social work.
Relation to Prior Work
Three lines of work motivate the gap.
The formal bargaining results supply the analytic machinery. Multi-Issue Negotiation with Deadlines shows that package deals solved backward from a deadline yield Pareto-optimal agreement. Bargaining in Bundle over Multiple Issues in Finite-Horizon Alternating-Offers Protocol shows that under additive utilities this reduces to a polynomial sequence of convex programs, with agreement at $t=1$. We reuse that computation. We do not claim people follow it.
Game-theoretic LLM: Agent Workflow for Negotiation Games prompts language models through explicit backward induction. This makes them far more rational in sequential games, but they remain exploitable by counterparts who do not play equilibrium. That paper evaluates only quantitatively and never codes transcripts. We turn the same computation toward explaining the counterparts.
Deal or No Deal? End-to-End Learning of Negotiation Dialogues motivates the question most sharply. Its reinforcement-learning agents with forward rollouts beat humans on score, yet agreement with humans fell from 76% to 57%. The paper does not explain why people walked away.
On the qualitative side, CaSiNo supplies a nine-strategy prosocial/proself taxonomy, with evidence that prosocial talk tracks satisfaction. It is descriptive and never relates strategies to equilibrium play. We use its taxonomy as the seed codebook.
LOGOS automates grounded-theory coding, meaning codes built up from the data rather than fixed in advance, and adds a train/test protocol for schemas. Its authors note that it captures only hierarchical relations, not sequential ones. We reuse its protocol and add the temporal anchor it lacks.
SocialRL finds that anticipating the counterpart's reaction predicts success far better than inferring their preferences. Its anticipation runs forward. We treat that finding as a hypothesis the codebook can test: the prediction is that high-loss turns cluster around misanticipated reactions rather than misread preferences.
Training Large Language Models for Reasoning through Reverse Curriculum Reinforcement Learning supplies the general intuition that working backward from terminal outcomes localizes credit. It applies this only to single-agent reasoning.
The mixed-methods studies establish why the qualitative track is needed at all. Kumari et al.'s consent study found that 83% of parents expected direct benefit despite valid consent forms. ECIR 2020 workshops found an attendance surge that masked collapsed networking. In both cases, headline numbers hid the mechanism.
Research Program
Analytic track. On Deal or No Deal, the dataset records both players' item values, and on CaSiNo it records both players' priorities. An LLM-based parser maps each utterance to an explicit offer, or to "no offer." Parser accuracy is audited on a hand-labeled sample. We impose a horizon $T$, equal to the maximum observed dialogue length, and run a sensitivity analysis over $T$ and the discount factor. This confronts the fixed-horizon limitation directly rather than hiding it.
For each turn $t$ with state $s_t$ (the standing offer and turn index), we compute the complete-information equilibrium continuation value $V^i_t(s_t)$ for the acting player $i$. We also compute a belief-based variant in which $i$'s knowledge of the partner's values is replaced by the empirical prior. The per-turn value loss is
$$\Delta^i_t = V^i_t(s_t) - V^i_{t+1}(s_{t+1}),$$
which is the value player $i$ forfeits by the move actually made. Comparing the two variants separates losses explained by missing information from losses that persist even with full information. The second kind are the interesting ones.
Qualitative track. Under a fixed budget of $k$ turns per dialogue, analysts code the top-$k$ loss turns. We use two analyst groups: trained human coders on a subset, and a LOGOS-style LLM pipeline at scale, with human–LLM agreement reported. Coding starts from the CaSiNo strategies and is open to new codes. We add a temporal relation type, "precedes-and-enables," so the schema can express sequences like undervalue-partner → reciprocal hardening → walkout. CaSiNo's post-dialogue satisfaction and partner-liking ratings serve as the quantitative survey stream for triangulation.
Integration. The coded dialogue state $c$ enters the terminal payoff as
$$u_i = x_i^\top v_i + \lambda^\top r(c),$$
where $x_i$ is $i$'s allocation, $v_i$ its private values, $r(c)$ a feature vector of coded relational states, and $\lambda$ weights estimated by maximum likelihood on training dialogues. Backward induction is then rerun under $u_i$. If $\lambda$ is near zero, the qualitative codes carry no decision-relevant content. That would be a clean negative result.
Baselines
The naive reference point is plain equilibrium plus choice noise. This is the complete-information backward-induction solution with a softmax over actions whose temperature is fit on training data. It asks whether noise alone explains human deviations. Beating it means a lower held-out negative log-likelihood of accept/reject decisions and final splits, with a paired-bootstrap interval excluding zero.
The strongest competitor for turn selection is an outcome-predictor attribution baseline. We fine-tune a language model on the same training dialogues to predict breakdown and satisfaction, then take its top-$k$ turns by attribution score. This is the most serious threat because it also finds "important" turns cheaply, with no game theory. It is trained on identical data with the same hyperparameter-search budget as the $\lambda$ fit. Beating it means codes induced from loss-selected turns reach a higher held-out breakdown AUC and satisfaction correlation at equal $k$, and score no worse on LOGOS parsimony and coverage.
For the coding component, the strongest reported reference is CaSiNo's multi-task strategy classifier at 68.3% overall F1 (CaSiNo). We use it both as a label source and as a fixed-codebook comparison: regress outcomes on its predicted labels across all turns. The reference for automated schema induction is LOGOS's 80.4% average expert alignment on multi-agent failure logs (LOGOS). That result is on a different corpus, and there is no reported result on Lacuna for LOGOS on negotiation dialogues. Held-out prediction of human breakdowns on Deal or No Deal has no reported result on Lacuna either. We therefore establish all baselines ourselves on one split.
Two ablations isolate the key ingredient. The first swaps loss-selected turns for random turns, then for all turns subsampled to the same $k$. The second uses the belief-based loss instead of the complete-information loss.
Why not just…?
Why not just let an outcome predictor find the pivotal turns? This is the real threat, and the experiments settle it head-on. A predictor finds where the outcome became predictable, which is often the final turn or an explicit rejection. Backward induction finds where value was given up relative to what the remaining horizon allowed. That can be several turns earlier, which is where an explanation would sit. If the attribution baseline matches loss-based selection on held-out prediction, the claim reduces to "any importance signal works," and we will report it that way.
Why not just code every turn with LOGOS? Coding every turn is feasible for an LLM but not for the human coders who validate it. Full coverage also dilutes the schema with routine turns: LOGOS itself shows that naive grouping captures surface similarity rather than mechanism. The equal-budget comparison makes this testable. If all-turn coding wins at matched human effort, the selection step is unnecessary.
Why not just fit a behavioral model such as noisy best response and skip qualitative work? That is exactly the naive baseline. The mixed-methods record in the file, including the consent study and the ML security survey, suggests aggregate fits hide the mechanism. Even if noise fits the choice data well, it cannot say why the 57% of agreements in Deal or No Deal occurred where the other 43% failed. The $\lambda$ test measures whether coded states add predictive content beyond noise.
Proposed Experiments
The first decisive check runs on Deal or No Deal and CaSiNo with fixed train and test splits. Loss-selected coding must beat random, all-turn, and attribution-selected coding at equal $k$ on held-out breakdown AUC and satisfaction correlation. Schemas are also scored on LOGOS's five dimensions: reusability, fitness, coverage, parsimony, and consistency.
The second check requires the augmented model to reduce held-out negative log-likelihood relative to noisy equilibrium. The gain must hold across the horizon and discount sensitivity grid. The SocialRL-derived hypothesis is tested by the share of high-loss turns coded as misanticipation versus misread preference.
A third, exploratory check applies the analysis to the 50 hardest Deal or No Deal cases used in Game-theoretic LLM. A backward-induction negotiator using $u_i$ faces an imitation-trained counterpart in place of humans. We measure whether agreement rate rises without the stubbornness collapse seen in Deal or No Deal. This is a secondary sanity check, not the main claim.
References
- Multi-Issue Negotiation with Deadlines
- Bargaining in Bundle over Multiple Issues in Finite-Horizon Alternating-Offers Protocol
- Game-theoretic LLM: Agent Workflow for Negotiation Games
- Deal or No Deal? End-to-End Learning of Negotiation Dialogues
- CaSiNo: A Corpus of Campsite Negotiation Dialogues for Automatic Negotiation Systems
- LOGOS: LLM-driven End-to-End Grounded Theory Development and Schema Induction for Qualitative Research
- From Passive Delegates to Strategic Negotiators: Reinforcing Social Reasoning in Small Language Models with SocialRL
- Training Large Language Models for Reasoning through Reverse Curriculum Reinforcement Learning
- Why parents consent to their children’s participation in genetic research: A study of parental decision making
- ECIR 2020 workshops: assessing the impact of going online
- Machine Learning Security in Industry: A Quantitative Survey