acceptodds
Under review as a conference paper at ICLR 2027

Social Inference is Not Enough: A Decision-Value Framework for Diagnosis and Intervention

Abstract

In strategic interactions, more accurate inference about an opponent does not necessarily translate into better decisions or higher payoffs. We develop a value-guided social inference framework to diagnose where this connection holds or breaks down. Building on value of information, the framework distinguishes three payoff comparisons relative to a prior-informed baseline: potential value, the improvement available from perfect knowledge; acquired value, the expected improvement enabled by an evidence procedure; and realized value, the informed policy's advantage in deployment, where evidence and opponent responses may differ from the source model. We characterize when remaining uncertainty has no decision value and connect payoff-relevant representations to cost-sensitive evidence selection and stopping. Controlled games isolate the absence of opportunities for better decisions, difficulty acquiring decision-relevant evidence, and loss of predicted gains under response mismatch. In two negotiation protocols with large language model agents, targeted interventions improve net return in the reported diagnostic comparisons by avoiding redundant identification, choosing cost-effective evidence, and verifying consequential but unreliable claims. Together, these results support a diagnostic account of how social inference contributes to interaction outcomes and illustrate how that account can guide what an agent learns, how it acts, and when it reassesses its model of the interaction.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Social Inference, Value of Information, Game Theory, Negotiation, Theory of Mind

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_xzcLOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This work argues that accurate social inference is not sufficient for good interactive decision making, and proposes a decision-value framework separating three quantities: potential value, the gain achievable under perfect knowledge of the opponent type; acquired value, the expected improvement from a particular evidence procedure; and realized value, the advantage of the resulting informed policy in deployment, potentially under evidence or opponent-response mismatch.

Critical strengths
  • Writing is easy to follow.
  • The formulation is mathematically well-defined.
  • The paper explicitly acknowledges the limitation that the framework is model-relative and that the experiments do not establish failure prevalence or general agent rankings.
Critical weaknesses
  1. Limited theoretical novelty. Potential and acquired value closely correspond to EVPI/EVSI, which the paper itself acknowledges, and several propositions appear to follow fairly directly from standard Bayesian decision theory.
  2. Each baseline is chosen to instantiate the specific failure mode being diagnosed, making Full's advantage somewhat expected.
  3. Some comparisons do not provide equal information access. In the final-offer inefficient-acquisition experiment, No language lacks the cheaper language-question channel available to Full, so the comparison partly measures channel availability rather than better decision making under identical information access.
  4. No strong common-access Bayesian baseline.
  5. Full performs better when the opponent lies but can perform worse than Naive VoI in honest cases because it pays unnecessary verification costs. The tradeoff that is underexplored.
Questions for authors
  1. What is the main theoretical contribution beyond standard EVPI/EVSI and Bayesian metareasoning?
  2. Can the authors provide a baseline with the same posterior, evidence channels, reliability estimates, payoff calculations, and numerical tools as Full, but without the proposed three-stage diagnostic framing?
  3. In the misplaced-trust setting, how does performance change as the probability of honest v.s. deceptive opponents varies? At what deception rate does verification become worthwhile overall?
  4. How much of Full's empirical improvement comes from the LLM reasoning component versus the deterministic Bayesian/value-computation module?
Reviewer_sup1OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a value-guided framework for social inference built on three payoff comparisons against a prior-informed baseline. Potential value is the gain from perfect knowledge of the opponent's type. Acquired value is the expected gain from a given evidence procedure. Realized value is the informed policy's advantage in a target environment where opponent responses may differ from the source model. The authors prove basic properties, including that potential value is zero exactly when a common optimal policy exists and that potential value decomposes into acquired value plus expected remaining potential. They run these quantities in strict ordinal 2×2 games. They then evaluate an LLM negotiation agent backed by an exact Bayesian value-of-information calculator against baselines matched to three failure modes: redundant identification, inefficient acquisition, and misplaced trust.

Critical strengths

The writing is good. The small analytical examples are good teaching material: oracle-action labels discarding payoff magnitudes (Example 1), entropy versus decision value, and complementary signals.

Critical weaknesses

The core quantities are standard. Potential value is the expected value of perfect information, acquired value is the expected value of sample information. The observation that information gain differs from decision value is also classical (Lindley vs. Blackwell). Applying value of information to negotiation is not new either, e.g., Mohammad & Nakadai (2019). The paper's contribution seems to be framing rather than on new results. The authors should state clearly what the framework predicts or explains that standard Bayesian decision analysis does not.

The negotiation is one shot game and closer to a single decision problem than to a negotiation.

Questions for authors

What does the framework predict or explain that standard Bayesian decision analysis and value of information do not?

Reviewer_u8YnOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a value-of-information framework for decomposing decision-making for social inference. The authors define potential value (value available under perfect information of the opponent), acquired value (expected gain from a new signal), and realized value (informed policy's gain at deployment with possible misspecification). These are defined following classical decision theory. The authors propose that these values can be used to distinguish between different failure modes in social inference decision-making (redundant identification, inefficient acquisition, misplaced trust). They numerically explore this framework in simple 2x2 games and in an LLM negotiation experiment. In the latter, an LLM agent is tasked with bilateral negotiation to determine an allocation of three items between itself and its opponent. The role of the LLM is to extract claims, estimate the opponent's reliability, and choose an action. These experiments illustrate the different signatures of various failure modes with respect to the authors' framework.

Critical strengths
  1. The framework of dissecting social inference using potential, acquired, and realized value is a clean use of decision-theoretic tools.
  2. The connection between the framework and corresponding failure modes is nice and provides an easy-to-understand taxonomy of errors.
Critical weaknesses
  1. There seems to be little to no technical novelty in the paper. All the formal content (Sections 3 and 4) is textbook decision theory given new names and framing. The appendix proofs are all standard. This leaves the "diagnostic framework" and the experiments as the remaining possible contributions.
  2. The diagnostic ability doesn't actually seem to be tested at all. The diagnostic groups in Section 6 are formed based on inference parameters in line with the analytical framework. The work does not explore the application of the framework to existing failures to triage their root cause.
  3. The experimental results seem to come by construction. This is natural for the numerical experiment of Section 5, but it seems to also be true of the LLM experiments in Section 6. The three baselines for comparison (prediction-first, no language, and naive VoI) are weaker than the "Full" agent following the authors' framework by assumption: they are given either a different objective, a different action space, or a hard-coded incorrect belief. This all but guarantees the results shown in the main body of the paper. This is made clear in Table 7 of the appendix as the baselines fall into the failure modes they are designed for but perform on par with the Full agent in other instances.
  4. The main body's description of the LLM experiments is minimal and rather vague. In order to really get a sense for the experiments and what is being tested, the reader needs to refer to the appendices. I would recommend providing a bit more detail up front.
Questions for authors
  1. Can you operationalize your framework so it can actually be used to diagnose an existing LLM negotiation (benchmark datasets are available) or other decision process? Without this, the "diagnostic" value seems unclear.
  2. What does the LLM contribute in your experimental setup? Could the same results not be achieved by just picking a simple heuristic for reliability estimate? This is reinforced by GPT and DeepSeek giving practically identical results.
  3. How would the Full agent (or more generally, your framework) apply when the environment (i.e. type space, acceptance rules, etc.) are not given? Can you speak to the practical implementation of your ideas?
  4. For the misplaced trust failure mode, Table 7 shows that Naive VoI outperforms the other agents on the honest core. This makes sense given that trusting an agent who happens to be honest is beneficial. Is there a reason you report only the strategic core in Table 1 rather than an average over both?
Reviewer_tWGdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes three-part decision-value framework in social inference. The paper develops formal properties of the proposed framework, illustrates its implications using controlled games, and evaluates targeted interventions in two LLM negotiation protocols. It addresses a meaningful problem and provides a clear framework.

Critical strengths

The paper proposes a clear diagnostic framework, offering a new perspective for social inference. The three social-inference value are theoretically defined, and the distinctions between them are also clearly articulated. In experimental research, the game-theoretic experiments provide a controlled setting for examining specific mechanisms, which I appreciate is a substantial amount of work. I think it has reference value for future game-theoretic benchmark design and interactive agent diagnosis.

Critical weaknesses

My primary confusion is whether the current three definitions of social-inference value provide an exhaustive classification that covers all the reasons why social inference might fail.

In my view, the paper's theoretical components largely build upon the classical VoI framework; it neither establishes a new algorithm directly applicable to a broad range of tasks nor provides a general benchmark to facilitate large-scale comparative studies. Therefore, I believe the contribution in terms of novelty is currently insufficient.

Also, I think the main text depends heavily on the appendix. Important experimental details and supplementary evidence are deferred to the appendix.

Questions for authors
  1. Are the three proposed definitions of decision value intended to form an exhaustive taxonomy, or are they only selected dimensions of a broader problem?
  2. What is the main theoretical contribution beyond the classical VoI framework? Or can you give some new insights that are not readily available from existing decision-theoretic analyses?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.