Source-linked AI summary
ORDDAR: Observation-Driven Reasoning for Distortion-Resilient Decision, Action, and Cognitive Recovery
Deblina Kar, Anant Nawalgaria, Shyamal Kumar Das Mandal
TL;DR
AI agents increasingly perform complex, long-horizon reasoning, but errors in intermediate states can propagate into unreliable decisions, while existing methods rarely localize and selectively repair those errors. ORDDAR models reasoning as cognitive state transitions, detects localized distortions, retrieves recovery complements, and repairs only affected states. Across mathematical, commonsense, multi-hop, and clinical benchmarks, it consistently outperforms evaluated reasoning baselines, particularly on CKDQA.
Problem
Intermediate reasoning errors can propagate through sequential agent reasoning and degrade final decisions, while existing approaches rarely localize and selectively repair faulty states.
Method
ORDDAR models reasoning as cognitive state transitions, detects localized distortions, retrieves complementary recovery patterns through a Cognitive Mirror, and selectively repairs affected transitions.
Results
ORDDAR consistently outperforms evaluated reasoning baselines across mathematical, commonsense, multi-hop, and clinical reasoning, particularly on CKDQA.
Takeaways & Limitations
Distortion detection, complement retrieval, and localized recovery each contribute to ORDDAR’s final performance, supporting recovery at the reasoning-transition level.
Takeaways & Limitations
Evaluation uses a private CKDQA dataset that is not publicly released, and test instances are excluded from the Cognitive Mirror before prediction to prevent contamination.
Abstract
from arXiv · showhide
AI agents increasingly perform long-term reasoning, planning, tool use, memory integration, and autonomous decision making, yet erroneous intermediate states can propagate and cause inconsistent decisions and unreliable outputs. Existing reasoning approaches mainly rely on iterative planning, self-reflection, augmented memory, or verification, but rarely localize and selectively repair faulty reasoning. We present ORDDAR (Observation-Driven Reasoning for Distortion-Resilient Decision, Action, and Cognitive Recovery), a reasoning framework that models reasoning as cognitive state transitions, detects localized distortions, retrieves related reasoning from prior experiences, and repairs only the affected states. ORDDAR therefore performs recovery at the local reasoning-transition level rather than regenerating the complete trajectory. Experiments across mathematical, commonsense, multi-hop, and clinical reasoning benchmarks demonstrate improved reasoning quality, recovery ability, and interpretability over multiple evaluated reasoning baselines.
1 INTRODUCTION
Agentic systems can perform complex autonomous workflows, but intermediate reasoning errors may propagate into inconsistent decisions. ORDDAR addresses this by detecting local distortions and selectively repairing affected reasoning states.
- Agentic AI combines perception, planning, tool use, memory, and autonomous workflows, yet complicated instructions and multiple constraints remain challenging.
- Existing approaches improve reasoning through interaction, search, or self-reflection, but intermediate errors can still propagate through later reasoning trajectories.
- ORDDAR models reasoning as cognitive state transitions and identifies localized distortions rather than reconstructing the entire reasoning path.
- The framework uses a Cognitive Mirror to combine distortion memory with complementary reasoning and repair only corrupted transitions.
2 RELATED WORK
Related work improves agent reasoning through interaction, search, reflection, and refinement, but longer or more elaborate reasoning paths can remain vulnerable to propagated intermediate errors. ORDDAR instead focuses on identifying and repairing corrupted reasoning paths.
- Recent agent frameworks support reasoning, planning, tool use, memory integration, and decision making, while benchmarks still report difficulties with instructions, constraints, and consistency.
- ReAct, Tree of Thoughts, Reflexion, and Self-Refine improve reasoning through interaction, path exploration, reflection, or refinement.
- Longer reasoning paths do not always improve reliability because intermediate errors may propagate and contaminate subsequent steps.
- ORDDAR specifically identifies reasoning distortions and repairs corrupted reasoning paths instead of generating an entire reasoning path.
3 METHODOLOGY
ORDDAR models reasoning as cognitive-state transitions, detects localized distortions, retrieves compatible recovery complements, and repairs only affected states while preserving valid trajectory components.
- 3.1 PROBLEM FORMULATION: ORDDAR represents sequential agent reasoning as observations transformed into recursively updated cognitive states and transitions toward prediction.The input query is converted into structured observations, which drive a trajectory of cognitive states through a reasoning transition function.
- 3.2 REASONING DISTORTION DETECTION AND SIGNATURE ENCODING: ORDDAR detects local reasoning failures by combining semantic-embedding displacement with cognitive-anchor displacement between consecutive states.The two normalized distances are combined into a joint distortion score, with a trajectory-adaptive threshold identifying distorted transitions.
- 3.2 REASONING DISTORTION DETECTION AND SIGNATURE ENCODING: Detected distortions are encoded with a Shift–Gain–Loss signature describing semantic shift, newly introduced anchors, and discarded anchors.The signature supports retrieval of a complementary recovery pattern without requiring an additional verifier LLM.
- 3.3 COGNITIVE MIRROR MEMORY: The Cognitive Mirror stores compact distortion signatures with associated recovery complements for localized, cross-task retrieval rather than complete reasoning trajectories.Its retrieval index is the distortion signature, while the associated complement provides a reusable recovery pattern.
- 3.4 COMPLEMENT RETRIEVAL: At runtime, ORDDAR retrieves the most compatible complement and applies it only to the affected reasoning state, preserving the remainder of the trajectory.The recovered trajectory replaces only the distorted state, avoiding unnecessary regeneration of valid reasoning states.
- 3.6 PREDICTION AND COGNITIVE MIRROR UPDATE: Only validated recovery patterns update the Cognitive Mirror, while test instances remain excluded before prediction to prevent evaluation-time contamination.The workflow otherwise remains task-agnostic, and trajectories are preserved when no compatible validated complement exists.
4 EXPERIMENTAL SETUP
The evaluation uses four reasoning benchmarks and standardized inference settings to compare methods fairly. Experiments cover mathematical, multi-hop, commonsense, and clinical reasoning.
- Benchmarks: Experiments use GSM8K, HotpotQA, StrategyQA, and CKDQA for mathematical, multi-hop, commonsense, and clinical reasoning, respectively.CKDQA was constructed from hospital cases following ethical clearance; the other datasets are publicly available.
- Evaluation settings: All methods use Qwen2.5-7B-Instruct with its standard tokenizer, chat template, deterministic greedy decoding, and a 512-token maximum.Decoding uses do sample=False.
- Evaluation settings: Each evaluation uses 100 test instances sampled with random seed 42.
5 RESULTS
Across the reported comparisons, ORDDAR achieves the highest performance while using localized distortion detection, complementary retrieval, and targeted recovery. Ablations and qualitative analyses indicate that these components support recovery across different reasoning tasks and error types.
- Overall performance: ORDDAR achieves the highest reported performance among CoT, Self-Refine, Reflexion, ToT, and ProCo.The comparison suggests that refinement effectiveness depends on task-specific features and intermediate errors.
- Overall performance: ORDDAR repairs affected reasoning states by identifying distortions and retrieving complementary reasoning patterns instead of reconstructing the whole reasoning path.
- Ablation study: Removing CM, DM, or RM consistently degrades performance, showing complementary roles for complement retrieval, distortion localization, and localized recovery.On CKDQA, removing CM, DM, and RM reduces performance from 94.38% to 81%, 78.05%, and 79.79%, respectively.
- Ablation study: DM removal causes the largest GSM8K drop, from 92% to 85%, highlighting the importance of accurate distortion localization.
- Distortion validation: Sensitivity analysis identifies a 60–20–20 weighting as best-performing, assigning 60% to transition similarity and 20% each to logical gain and loss.Ground-truth detection is evaluated across GSM8K, HotpotQA, StrategyQA, and Clinical-CKD using precision, recall, and F1.
- Qualitative analysis: In a Clinical-CKD case, ORDDAR localizes a distortion at 0.874, retrieves a complement at 0.940, and repairs only the erroneous transition.The example corrects an initial G3 assignment to G4 for eGFR 16.6 mL/min/1.73m2.
6 CONCLUSION
ORDDAR detects localized reasoning distortions, retrieves complementary recovery patterns, and repairs affected states across several reasoning benchmarks. The study also reports component ablations, retrieval analysis, ethical protections for clinical data, and implementation details for reproducibility.
- 6 CONCLUSION: ORDDAR consistently outperformed evaluated reasoning baselines across mathematical, commonsense, multi-hop, and clinical reasoning tasks.The strongest reported performance was on CKDQA.
- 6 CONCLUSION: Ablation results show that distortion detection, complement retrieval, and recovery each contribute to final performance.
- 6 CONCLUSION: Retrieval analysis shows that distortion signatures can identify relevant recovery patterns.
- 6 CONCLUSION: Handling higher-order multi-hop dependencies remains an important direction for future research.Clinical recovery complements currently incorporate locally provided KDIGO knowledge.
- 6 CONCLUSION: The clinical study used ethically approved data, protected patient privacy, and did not publicly release the private CKDQA dataset.
- 6 CONCLUSION: The released implementation covers distortion detection, Cognitive Mirror retrieval, localized recovery, and evaluations across four datasets.