Wrong but Useful: Trajectory Value Beyond Answer Correctness in Multi-Agent Messages

Abstract:Multi-agent reasoning systems often use agreement, confidence, or automated scores to decide which messages should shape a final answer. Such filtering assumes that a message likely to be correct is also worth keeping. Yet a wrong answer can contain a useful decomposition, constraint, or scientific principle. We test this distinction with Diverse Hypothesis Deliberation (DHD), a controlled measurement protocol that caches five independently generated messages and replays the same downstream solver, called the integrator, with each message available or hidden. The replay comparison measures a message's trajectory value: whether making the message available helps or harms subsequent reasoning. Across five mathematics and science benchmarks and two openly available model families, gpt-oss-120b and gemma-4-31B-it, wrong-helpful messages appear in every benchmark-model combination. Among wrong-answer messages that change final correctness, more than four in ten changes are helpful in each model. Controlled repeats show that the number of repeatable message effects is unlikely to arise from replay variation alone (p=0.0002). A focused intervention on repeatable wrong-helpful messages finds that the complete message works best, while retaining its reasoning preserves more success than retaining only its answer; the source of the complete-message advantage remains open. Within the same problem, repeated trajectory-value evidence also identifies a better keep-or-remove choice than answer correctness alone. Answer correctness is therefore informative but does not determine trajectory value. DHD measures this missing property and produces reusable labels for learning when agents should listen.

24 pages, 9 figures. Includes an appendix and an ancillary reproducibility artifact

Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG)

arXiv:2608.14375 [cs.AI]

(or arXiv:2608.14375v1 [cs.AI] for this version)

https://doi.org/10.48550/arXiv.2608.14375 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

anonymous_reproducibility/ARTIFACT_MANIFEST.json

anonymous_reproducibility/DATA.md

anonymous_reproducibility/ENVIRONMENT.md

anonymous_reproducibility/README.md

anonymous_reproducibility/REPRODUCE.md

anonymous_reproducibility/THIRD_PARTY_NOTICES.md

anonymous_reproducibility/analysis/analyze_repeated_effect_by_correctness.py

anonymous_reproducibility/analysis/problem_blocked_replay_multiplicity.py

anonymous_reproducibility/analysis/run_majority_diagnostic.py

anonymous_reproducibility/analysis/run_offline_review_closure.py

anonymous_reproducibility/analysis/run_proposal_answer_audit.py

anonymous_reproducibility/analysis/test_problem_blocked_replay_multiplicity.py

anonymous_reproducibility/configs/paper/dhd_config.yaml

anonymous_reproducibility/configs/paper/dhd_config_secondary_actor.yaml

anonymous_reproducibility/configs/smoke/smoke_config.yaml

anonymous_reproducibility/data/derived/component_masking_pilot_summary.json

anonymous_reproducibility/data/derived/correctness_oracle_regret_decomposition.csv

anonymous_reproducibility/data/derived/gemma_repeated_benchmark_summary.json

anonymous_reproducibility/data/derived/integrator_uptake_audit.csv

anonymous_reproducibility/data/derived/majority_diagnostic.csv

anonymous_reproducibility/data/derived/majority_diagnostic.json

anonymous_reproducibility/data/derived/offline_policy_bootstrap.csv

anonymous_reproducibility/data/derived/offline_policy_by_benchmark.csv

anonymous_reproducibility/data/derived/offline_review_closure.json

anonymous_reproducibility/data/derived/oss_proposal_answer_audit.json

anonymous_reproducibility/data/derived/problem_blocked_replay_multiplicity.json

anonymous_reproducibility/data/derived/proposal_answer_coverage.csv

anonymous_reproducibility/data/derived/repeated_effect_by_correctness.json

anonymous_reproducibility/data/derived/signal_labels/matched_sensitivity_integrator_gemma_to_oss.csv

anonymous_reproducibility/data/derived/signal_labels/matched_sensitivity_order_gemma.csv

anonymous_reproducibility/data/derived/signal_labels/matched_sensitivity_order_oss.csv

anonymous_reproducibility/data/derived/signal_labels/value_aware_one_removal_per_fold.csv

anonymous_reproducibility/data/fixtures/smoke_pool.jsonl

anonymous_reproducibility/data/manifests/expected_signal_tables.json

anonymous_reproducibility/data/manifests/expected_summaries.json

anonymous_reproducibility/figures/benchmark_properties_source.csv

anonymous_reproducibility/figures/evaluator_agreement_source.csv

anonymous_reproducibility/figures/fig_correctness_flip_direction_source.csv

anonymous_reproducibility/figures/fig_cross_benchmark_trajectory_value_source.csv

anonymous_reproducibility/figures/fig_generation_integration_gap_gemma_source.csv

anonymous_reproducibility/figures/fig_generation_integration_gap_source.csv

anonymous_reproducibility/figures/fig_loo_by_k_appendix_source.csv

anonymous_reproducibility/figures/make_correctness_flip_direction.py

anonymous_reproducibility/figures/make_cross_benchmark_trajectory_value.py

anonymous_reproducibility/figures/make_evaluator_agreement.py

anonymous_reproducibility/figures/make_generation_integration_gap.py

anonymous_reproducibility/figures/make_loo_by_k_appendix.py

anonymous_reproducibility/pyproject.toml

anonymous_reproducibility/requirements.lock

anonymous_reproducibility/scripts/recompute_signal_tables.py

anonymous_reproducibility/scripts/reproduce_analysis.py

anonymous_reproducibility/scripts/reproduce_analysis.sh

anonymous_reproducibility/scripts/run_smoke_test.py

anonymous_reproducibility/scripts/run_smoke_test.sh

anonymous_reproducibility/scripts/validate_inputs.py

anonymous_reproducibility/src/dhd/__init__.py

anonymous_reproducibility/src/dhd/labels.py

anonymous_reproducibility/src/dhd/model_api.py

anonymous_reproducibility/src/dhd/protocol.py

anonymous_reproducibility/src/dhd/replay.py

anonymous_reproducibility/tests/conftest.py

anonymous_reproducibility/tests/test_artifact_anonymity.py

anonymous_reproducibility/tests/test_labels.py

anonymous_reproducibility/tests/test_model_api.py

anonymous_reproducibility/tests/test_protocol.py

anonymous_reproducibility/tests/test_replay.py

anonymous_reproducibility/tests/test_reproduce_analysis.py

anonymous_reproducibility/tests/test_signal_tables.py

(65 additional files not shown) You must enabled JavaScript to view entire file list.

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.