Tag
This paper formalizes semantic coordination-strategy unlearning in cooperative MARL and proposes CARVE, which uses causal coordination attribution, retain-aware subspace updates, counterfactual re-coordination, and adversarial auditing to remove obsolete conventions while preserving useful cooperation, achieving 91.4% reduction in convention elicitation with only 5.9% median utility loss across five benchmarks.
This paper proposes using controlled synthetic pretraining with the O'PRIOR task generator to make training-data attribution experimentally verifiable for tabular foundation models, showing that top-attributed task removal measurably degrades held-out performance and revealing that provenance association and interventional faithfulness can diverge.
Introduces CAS, a causal attribution score for local and global explainable AI that separates predictive importance from causal effect heterogeneity, demonstrated on benchmarks and empirical datasets.
Introduces TelemetrySuffBench, a benchmark for evaluating whether agent telemetry is sufficient for failure-origin diagnosis. Finds that full telemetry enables high origin-step accuracy for some models, but coarse/OpenTelemetry-compatible views create a strong detection-localization gap, and several models struggle with safe abstention on ambiguous traces.
This paper identifies that naive skill accumulation in LLM agents can cause performance regressions, as skills beneficial for some tasks hurt others. The authors propose Assay, a framework that measures per-skill causal contributions and applies per-task masking, achieving state-of-the-art results on AppWorld and τ-bench without weight updates.
This paper investigates the developmental conditions under which a minimal predictive neural system (a 192-dimensional GRU) can distinguish self-caused changes from world-caused changes, identifying four necessary conditions for agency and introducing a metric called agency gain.
MemAudit is a post-hoc auditing framework for memory-augmented LLM agents that identifies poisoned memories by combining counterfactual influence scores and structural anomaly detection, reducing attack success rates from over 70% to 0% in realistic scenarios.