When Is Enough Not Enough? Illusory Completion in Search Agents

Abstract: Recent search agents leverage multi-turn reasoning and search tools to achieve strong performance on multi-hop and long-horizon benchmarks. Yet it remains unclear whether they reliably reason across all requirements by tracking, verifying, and maintaining multiple conditions in these questions. We study this capability under multi-constraint problems, where valid answers must satisfy several constraints simultaneously. We find that illusory completion frequently occurs, wherein agents believe tasks are complete despite unresolved or violated constraints, leading to underverified answers. To diagnose this behavior, we introduce the Epistemic Ledger, an evaluation framework that tracks evidential support and agents' beliefs for each constraint throughout multi-turn reasoning. Our analysis reveals four recurring failure patterns: bare assertions, overlooked refutations, stagnation, and premature exit. Motivated by these findings, we examine whether explicit constraint-state tracking during execution mitigates these failures via LiveLedger, an inference-time tracker. This simple intervention consistently improves performance, substantially reducing underverified answers (by up to 26.5%) and improving overall accuracy (by up to 11.6%) on multi-constraint problems.

• The paper introduces the Epistemic Ledger, a trajectory-level method that tracks evidence and agent belief for each candidate–constraint pair, achieving 93% human-validated detection accuracy.
• Experiments on 215 multi-constraint problems show underverification remains widespread—even correct answers can lack support—with leading agents exceeding 50% underverified answers and four recurring failure mechanisms identified.
• LiveLedger exposes constraint states during inference, increasing accuracy by up to 11.6 points and reducing underverification by up to 26.5 points, outperforming simply forcing agents to search longer.

This paper investigates a failure mode of LLM-based search agents that final-answer evaluation obscures: illusory completion, the epistemic state in which an agent terminates and reports an answer believing all constraints of a multi-constraint problem (MCP) are satisfied when at least one remains unverified or violated. The authors introduce the Epistemic Ledger, a trajectory-level evaluation framework that jointly tracks evidential support and agent belief per candidate–constraint pair, diagnose four recurring failure mechanisms, and propose LiveLedger, an inference-time constraint tracker that measurably reduces underverification and improves accuracy (2602.07549).

Motivation and problem formulation

Search agents built on the ReAct paradigm interleave reasoning traces $\tau_t$, actions $a_t$ (search, browse, answer), and observations $o_t$, producing a trajectory $\mathcal{H}_T$ whose final action is the predicted answer. Standard multi-hop and long-horizon benchmarks (MuSiQue, FRAMES, BrowseComp, WebWalkerQA, LiveDRBench) grade only the final answer, so an agent that reaches a correct answer via partial evidence is indistinguishable from one that verified every requirement. The authors argue this hides epistemic failures even under correct answers.

They isolate this behavior on multi-constraint problems, where a valid answer must jointly satisfy an explicit constraint set $\mathcal{C}=\{C_1,\dots,C_n\}$. An answer is underverified if the agent commits to it while some $C_i$ is violated or unsupported by retrieved evidence. The key definitional move is separating what the evidence supports from what the agent believes — a divergence between the two constitutes illusory completion.

The Epistemic Ledger

The Epistemic Ledger is maintained by an evaluator model (gpt-oss-120b) that processes the trajectory incrementally. For each candidate $k$ and constraint $C_i$, it records:

• Evidential support $E(k, C_i) \in \{Satisfied, Refuted, Unknown\}$, derived only from observations;
• Agent belief $B(k, C_i) \in \{Affirm, Deny, Unaddress\}$, derived only from the agent's reasoning traces;
• Candidate status $a_t$0, capturing the agent's commitment.

Updates proceed in two stages per step: evidence states from $a_t$1, then beliefs and status from the subsequent reasoning $a_t$2. An answer is flagged underverified when its candidate is Active and some constraint has $a_t$3. Human validation of this detection is strong: 93% overall accuracy, 100% majority-vote accuracy, and substantial inter-annotator agreement ($a_t$4).

Analyzing joint $a_t$5 dynamics yields four failure mechanisms:

1. Bare assertion: $a_t$6 with $a_t$7 — the agent claims satisfaction without evidence.
2. Overlooked refutation: $a_t$8 but $a_t$9 — disconfirming evidence is ignored.
3. Stagnation: ledger state unchanged over the last $o_t$0 steps while a constraint remains Unknown — circular search terminating without resolution.
4. Premature exit: termination with $o_t$1 and $o_t$2 — a constraint never addressed.

These mechanisms may co-occur in a single trajectory, and the framework's value lies in making them visible where final-answer correctness cannot.

Experimental setup

The authors assemble 215 MCP instances from five benchmarks (BrowseComp, DeepSearchQA, FRAMES, LiveDRBench, WebWalkerQA) by decomposing questions into DAGs and selecting those with width $o_t$3 and depth $o_t$4, i.e., questions requiring three or more independent constraints merged into a single entity. They evaluate eight search agents spanning RL-trained (Search-R1, ASearcher, DR-Tulu, WebExplorer, Tongyi DeepResearch), SFT (HybridDeepSearcher), and prompt-based (Search-o1, ReAct with gpt-oss-20b/120b) systems, using accuracy (judged by gpt-5) and underverified answer rate (UAR) as metrics. LiveLedger is applied to the strongest RL-trained and prompt-based baselines, TongyiDR and ReAct.

Main findings

Illusory completion is pervasive, even among strong agents. Accuracy ranges from 7.0 (ASearcher) to 56.7 (TongyiDR), but even the strongest systems exhibit UAR above 50% (TongyiDR 52.1, WebExplorer 72.6). Notably, underverification persists within correct answers: 19.1% of TongyiDR's answers are correct yet underverified. This directly supports the paper's central claim that final-answer correctness does not reflect procedural soundness.

System	Acc (%)	UAR (%)
ASearcher	7.0	94.9
Search-R1	13.0	93.5
DR-Tulu	17.2	90.2
HDS	16.3	86.5
Search-o1 (20B)	24.2	76.3
ReAct (20B)	36.3	62.8
WebExplorer	36.3	72.6
Search-o1 (120B)	34.4	65.6
ReAct (120B)	39.1	76.3
ReAct + TTS (120B)	41.3	52.9
TongyiDR	56.7	52.1
TongyiDR + LiveLedger	58.9	51.4
ReAct + LiveLedger (120B)	50.7	49.8

RL training and model scaling redistribute rather than eliminate failure modes. This is one of the paper's more contrarian claims. As RL training intensifies (Search-R1 → WebExplorer → TongyiDR), bare assertion and premature exit shrink, but overlooked refutation worsens and stagnation persists; bare assertion is 20% for TongyiDR versus 7% for ReAct-20B. Scaling from 20B to 120B reduces premature exit for Search-o1 (61% → 47%) but increases bare assertion (20% → 27%), and for ReAct it worsens all four modes. Longer, more capable trajectories touch more constraints, but answers are still finalized without sufficient evidence.

LiveLedger improves accuracy and reduces underverification. LiveLedger updates the ledger online after each turn and injects it into the agent's context (concatenated into the observation for post-trained agents to avoid destabilizing their learned behavior). Gains: +2.2 accuracy for TongyiDR-L-20B, +5.1 for ReAct-L-20B, and +11.6 for ReAct-L-120B, with UAR reductions up to 26.5 points (ReAct-120B 76.3 → 49.8). LiveLedger most strongly mitigates premature exit (ReAct-120B: 60% → 35%) and stagnation (56% → 41%).

LiveLedger outperforms naive test-time scaling. A ReAct-TTS variant (s1-style suppression of answer generation, forcing additional search turns) reduces UAR to 52.9 and accuracy to 41.3 — but leaves a 9.4-point accuracy gap to LiveLedger, while adding roughly 26.7 extra turns. The breakdown shows TTS mainly verifies already-correct answers without correcting errors, whereas LiveLedger increases the Extent of Candidate Exploration (ECE), a new metric of distinct candidates explored per turn — nearly doubling exploration for ReAct-L-20B. Explicit constraint states thus steer exploration rather than merely prolonging search; LiveLedger also reduces trajectory length, improving efficiency.

Limitations and open questions

The paper is candid about several constraints. The evaluation set is small (215 instances, filtered to depth-1 DAGs), and LiveLedger's effectiveness is backbone-dependent: with a 20B backbone it increases bare assertion and overlooked refutation, and stagnation worsens for ReAct-20B (34% → 42%), suggesting smaller models struggle to digest ledger updates. The ledger itself is model-based (gpt-oss-120b); manual review of verified-but-incorrect cases found 2 of 37 were evaluation misclassifications caused by truncated web snippets, and 24 were correct answers missing from ground truth — indicating both residual framework error and gold-annotation noise. LiveLedger also adds inference-time overhead via an external updater. The mechanism-level diagnosis relies on the evaluator correctly attributing beliefs to reasoning traces, an assumption validated only on a 30-question sample. Open questions include whether ledger-aware behavior can be internalized via post-training rather than imposed at inference, and whether the four failure mechanisms generalize beyond depth-1 multi-constraint QA.

Conclusion

This work reframes search-agent evaluation around constraint-level verification rather than final-answer correctness. The Epistemic Ledger provides a reusable diagnostic that exposes illusory completion and decomposes it into four actionable mechanisms; LiveLedger demonstrates that simply surfacing constraint states at inference time yields consistent accuracy gains (up to +11.6) and underverification reductions (up to 26.5 points) that test-time scaling cannot match. The finding that scaling and RL redistribute failure modes rather than removing them is the paper's most consequential claim for anyone training or deploying long-horizon search agents.