AI Research Agents Fail at Thinking, Not Writing: 45 Failure Patterns Across 100 Tasks

The dominant assumption about AI research agents is that language quality is the ceiling: get the generation fluent enough and scientific output follows. Across 800 agent trajectories on 100 real frontier research tasks, that assumption inverts. Agents write well. They fail to think about whether what they wrote is actually right.

AutoResearchEval covers the full research lifecycle, from ideation through retrieval, execution, analysis, writing, and review, across 7 scientific domains. Eight harness-model combinations each ran 100 tasks, producing trajectories annotated at the process level, not just scored at the output level. That distinction matters. Aggregate scores hide the mechanism of failure; trajectory-level inspection reveals it. The result is ARFT, the AutoResearch Failure Taxonomy: 45 empirically-grounded failure patterns organized from the ground up, not imposed top-down.

The 45 patterns converge on a single structural deficit. Think of it as a missing inner loop. A skilled researcher who finishes an experiment checks whether the result actually matches the hypothesis, whether the method they chose was the right one, and whether an anomaly in the data deserves a second look before writing it up. Current agents skip that loop entirely. They produce output, but they do not audit it. They do not catch when their retrieved evidence contradicts their stated conclusion. They do not revise when an intermediate result fails to hold. The failure mode is not hallucination in the conventional sense; it is the absence of the corrective mechanism that would catch the hallucination before it propagates into the final artifact.

What makes this finding structurally significant is where the deficit lives. The same 45 patterns appear across all 8 harness-model combinations, including the strongest models tested. Swapping the scaffold does not fix it. Swapping the model does not fix it. The problem is located at the model level, not in any particular orchestration layer. That rules out a class of quick fixes that practitioners might reach for first.

45 failure patterns, 800 trajectories, 7 domains, and the failure signature is consistent across every combination tested. For teams building AI co-scientist pipelines, the takeaway is direct: debugging your orchestration layer will not close the gap, because the gap is not in the orchestration.

We're thinking: We find the failure localization more consequential than the taxonomy itself. The field has spent considerable effort on scaffold design, tool routing, and memory architecture, treating the model as a fixed substrate to be orchestrated around. ARFT suggests that framing is wrong at the root. If the metacognitive deficit is model-level, then the fix has to be model-level too: training regimes that reward self-audit, verification steps that are scored rather than optional, and evaluation setups that penalize confident-but-wrong outputs rather than just rewarding final-answer accuracy. The 45 patterns are a diagnostic instrument. The harder question they expose is whether reinforcement learning from outcome-based rewards can actually install a metacognitive loop, or whether that requires a fundamentally different training objective.

Key takeaways:

• Current AutoResearch agents lack a metacognitive loop: they produce outputs but do not check those outputs against intermediate findings, revise when results fail to hold, or question whether their chosen path was sound.
• Across 800 trajectories and 8 harness-model combinations on 100 frontier research tasks, the same 45 failure patterns recur consistently, with the deficit localized to the model level rather than any scaffold; the benchmark and taxonomy are publicly released, though orchestration-level interventions remain untested.
• Teams building AI co-scientist or autonomous research pipelines should audit their evaluation setup first: if you are only scoring final outputs, you are blind to the process-level failures that ARFT reveals, and fixing scaffold routing will not address a model-level metacognition gap.

Source: AutoResearchEval: End-to-End Diagnostic Evaluation on 100 Real-World Frontier Research Tasks