The Self-Correction Illusion: LLMs Correct Others but Not Themselves

Paper 2 likely has higher impact: it delivers a programmable, deployable system that turns sparse-attention research into real serving-speed gains on large, modern models/GPUs, with clear real-world applicability and cross-field relevance (systems + ML + agentic optimization). The reported throughput improvements on production-scale models suggest broad adoption potential and timeliness as context lengths grow. Paper 1 is novel and insightful for prompt/agent evaluation and safety, but its impact is more scoped to chat-template behavior and prompting interventions, with less immediate infrastructural leverage.

Paper 1 challenges a fundamental assumption about LLM capabilities, revealing that the inability to self-correct is a structural artifact rather than a cognitive deficit. This novel finding has profound implications for LLM evaluation, prompting, and agent design across all domains, offering a highly generalizable intervention. While Paper 2 addresses a critical, high-stakes issue in healthcare with a valuable dataset, its impact is confined to the specific niche of medical deepfake detection. Paper 1's broader relevance to the rapidly growing field of LLM reasoning gives it a higher potential scientific impact.

Paper 2 identifies a fundamental mechanistic insight about LLM self-correction failures—that they stem from chat-template role labels rather than cognitive deficits—with rigorous experimental design (SHA-256 verified identical claims, 13 model-domain cells, statistical significance). This finding has immediate practical applications (training-free prompt interventions) and broad implications for all LLM deployment. Paper 1, while creative in applying Navya-Nyaya logic, is limited by a tiny dataset (55 problems), a single evaluation showing format adherence issues, and unclear generalization beyond the specific reasoning tasks tested.

Paper 1 offers a more technically novel and broadly enabling approach: compressing explicit reasoning into latent states and coupling it with a generative world-model objective for anticipating future UI states. This can materially improve efficiency and reliability of real-world mobile/UI agents, with quantitative gains and clear deployment relevance. While Paper 2 is timely and rigorous, its main contribution is a diagnostic/measurement finding plus a prompt-structure workaround; impactful for evaluation and prompting practice but narrower in downstream capability advances than MIRAGE’s method and application scope.

Paper 1 is more novel and broadly impactful: it isolates a surprising, causal “role-label” mechanism behind self-correction failures using byte-identical claims and shows large, statistically strong effects across many model families/domains. It yields an immediate, training-free intervention with implications for evaluation methodology, agent design, alignment, and reproducibility—relevant across most LLM applications. Paper 2 targets an important domain, but its approach (curriculum learning + multi-model reranking/selection) is more incremental, evaluated on a single dataset with limited rigor details and narrower cross-field impact.

Paper 1 reveals a fundamental and broadly impactful insight about LLM self-correction—a widely studied problem—showing it's a chat-template artifact rather than a capability deficit. This finding has immediate implications for the entire LLM reasoning and alignment community, offers a training-free fix, and challenges prevailing assumptions. Paper 2 addresses a more niche problem (RDB foundation models) with solid but incremental contributions. Paper 1's breadth of impact across LLM research, its mechanistic clarity, and its timeliness given the current focus on LLM reasoning give it higher potential scientific impact.

Paper 1 reveals a fundamental and surprising mechanism underlying LLM self-correction failures—showing it's a chat-template artifact rather than a capability deficit. This has broad implications across all LLM applications involving reasoning and self-correction, offers a training-free intervention, and challenges widespread assumptions in the field. Paper 2, while practically useful for code localization efficiency, addresses a narrower engineering optimization problem. Paper 1's finding is more novel, more broadly impactful, and likely to influence how the community designs prompting strategies and understands LLM behavior.

Paper 2 addresses a major, widely debated issue in LLM research—self-correction—and reveals it as a chat-template artifact rather than a capability deficit. This paradigm-shifting insight, combined with a highly practical zero-training intervention, offers immediate, broad applicability across agentic workflows. While Paper 1 provides valuable methodological improvements for evaluating reasoning metrics, Paper 2's findings fundamentally alter how researchers and practitioners understand and implement LLM reflection, likely leading to wider and more rapid adoption.

Paper 2 reveals a fundamental mechanistic insight about LLM behavior—that self-correction failures are chat-template artifacts rather than capability deficits—with broad implications across all LLM applications. It provides a training-free, model-agnostic intervention applicable immediately. The finding is surprising, rigorously controlled (SHA-256 verified identical claims, 13 model-domain cells, statistical significance), and affects the large community working on LLM reasoning and self-correction. Paper 1, while useful for pandemic forecasting, is more narrowly applied and incremental in its agent-memory architecture contribution.

Paper 2 offers a more novel and broadly relevant causal finding: self-correction failures largely arise from chat-template role labeling, isolated via byte-identical claims and verified controls. It is timely for agentic LLM design, has immediate real-world application (a training-free prompt-structure intervention), and generalizes across multiple model families and domains with strong statistical evidence. Paper 1 is insightful and methodologically solid but is narrower (DSL program evolution setting) and its impact is more specialized to LLM-guided genetic programming, whereas Paper 2 affects evaluation protocols, safety, reliability, and interface design across many LLM applications.

Paper 2 likely has higher impact because it targets a broadly relevant, timely safety failure mode—inference-time adversarial perturbations across the whole generation trajectory—and proposes a general training remedy (trajectory-based alignment) with clear real-world implications for deployed LLM robustness. Its claims connect to security, alignment, and sequence modeling, with potential downstream influence on evaluation protocols and training curricula. Paper 1 is novel and actionable (role-label artifact), but its mechanism may be more template-specific and narrower in scope compared to trajectory-level robustness that applies across interfaces and deployment settings.

Paper 2 is more novel and timely, identifying a causal, role-label-driven artifact in LLM self-correction with a tight controlled design (byte-identical claims, multi-model/multi-domain, strong statistics). It has immediate real-world applications for prompt design, evaluation protocols, and safety/alignment (reducing error persistence without retraining). Its implications span NLP, HCI, agent design, and benchmarking. Paper 1 addresses an important CV problem and proposes a reasonable architectural tweak, but the conceptual advance and cross-field breadth are narrower and impact is likely more incremental.

Paper 2 likely has higher impact: it introduces a new benchmark (ToolMaze) targeting a timely, real-world-critical gap—tool failures and dynamic replanning in LLM agents—yielding broadly applicable metrics (e.g., PRR) and insights that generalize across models, tools, and deployment settings. Its focus on robustness/fault tolerance affects many applied domains and can become a standard evaluation suite. Paper 1 is novel and actionable, but is more narrowly scoped to chat-template role artifacts and may be partially mitigated by interface changes, limiting breadth and longevity.

Paper 1 identifies a fundamental and surprising mechanism—self-correction failure in LLMs is a chat-template artifact, not a capability deficit—which challenges prevailing assumptions about LLM reasoning limitations. It offers a training-free, immediately deployable intervention with large effect sizes across multiple model families. This insight has broad implications for LLM agent design, RLHF training, and prompt engineering. Paper 2 provides a valuable systems characterization of agent memory but is more incremental (taxonomy, profiling, benchmarking) without a similarly transformative finding. Paper 1's causal mechanistic insight is more likely to reshape research directions.

Paper 2 offers a highly novel, clean causal identification of a widely observed phenomenon (self-correction failure) as a chat-template role-label artifact, using byte-identical claims and broad cross-model/domain evidence. Its immediate, training-free intervention has clear real-world applicability for agent prompting and evaluation, making it timely and actionable. While Paper 1 is rigorous and important for faithfulness in schema-guided reasoning, its contribution is more incremental within existing faithfulness/tool-use debates and the remediation leans on external tools or preference optimization. Paper 2’s finding likely impacts prompting, safety, and agent design broadly.

Paper 2 identifies a specific, mechanistically grounded phenomenon (self-correction failure as a chat-template artifact rather than a capability deficit) with immediate practical implications. It offers rigorous experimental methodology (SHA-256 verified byte-identical claims, 13 model-domain cells, statistical significance), actionable interventions requiring no retraining, and broad applicability across LLM deployments. Paper 1 addresses an important philosophical question but is more speculative, relying on worked case studies rather than empirical validation, and targets a narrower audience. Paper 2's finding is more likely to influence widespread LLM engineering practices and spawn follow-up research.

Paper 1 resolves a major ongoing debate in LLM research regarding self-correction capabilities, demonstrating that failures are driven by chat-template artifacts rather than cognitive deficits. This is a highly novel, paradigm-shifting insight with immediate, broadly applicable real-world implications for prompt design and agentic workflows. While Paper 2 offers a valuable benchmark for optimization-like reasoning, Paper 1's fundamental mechanistic discovery about model behavior gives it broader cross-disciplinary impact and higher immediate relevance.

Paper 1 challenges a fundamental assumption about LLM reasoning, revealing that self-correction failures are chat-template artifacts rather than cognitive deficits. This paradigm-shifting insight has profound implications across all LLM research, agent design, and alignment. While Paper 2 achieves impressive state-of-the-art results in formal theorem proving, its impact is largely confined to the AI-for-math subfield. Paper 1 offers broader applicability, rigorous mechanistic insights, and an immediate, training-free intervention for the broader AI community.

Paper 2 likely has higher scientific impact: it introduces a general formal framework and metrics for appropriate reliance with set-valued AI advice, a timely and growing practice for uncertainty communication. The contribution is broadly applicable across human-AI interaction, decision science, and ML evaluation, and can standardize measurement in many experimental paradigms (classification/regression, sequential settings). Paper 1 is novel and practically useful for LLM prompting, but its core claim is tied to chat-template artifacts and may be less generalizable across future model interfaces; its impact is strong but narrower and potentially more transient.