Decomposing how prompting steers behavior

Paper 2 offers foundational insights into mechanistic interpretability, uncovering the geometric transformations underlying prompting in LLMs and VLMs. While Paper 1 introduces a highly relevant benchmark for evaluating recursive self-improvement, Paper 2 provides deep causal explanations of model internals across multiple architectures and modalities. This fundamental approach to understanding the 'how' and 'why' of neural network behavior is likely to drive broader, more enduring scientific advancements across the field of AI.

Paper 1 addresses fundamental mechanistic interpretability by exploring how prompts alter internal representational geometry in foundation models. This theoretical contribution has broad implications for understanding and steering LLMs and VLMs across various tasks. While Paper 2 offers a valuable and practical memory architecture, Paper 1's insights into the core mechanics of prompting are likely to have a more profound and widespread impact on future AI research.

Paper 1 introduces a novel, general-purpose geometric decomposition framework for understanding how prompting reshapes internal representations in LLMs/VLMs. It provides fundamental mechanistic insights applicable across all prompted models, with rigorous causal testing methodology. Its breadth—spanning multiple model architectures, modalities, and tasks—gives it wide relevance to the interpretability and alignment communities. Paper 2 is a solid engineering contribution for biomedical agents but is more domain-specific and incremental (combining MCP servers with graph-based planning). Paper 1's foundational insights into prompt mechanisms have broader and longer-lasting scientific impact.

Paper 2 likely has higher scientific impact due to broader cross-model and cross-modality relevance and a more fundamental mechanistic contribution: a causal, layered geometric decomposition of how prompts steer internal representations and behavior. This advances interpretability and prompt engineering across LLMs and VLMs, with potential implications for controllability, safety, and neuroscience-style analysis. Paper 1 addresses an important RL-agent failure mode with a practical mitigation, but it is narrower in scope (agentic RL training setups) and the core method (advantage reweighting with critiques) is comparatively incremental.

Paper 2 tackles a fundamental theoretical question in AI: how prompting mechanically alters internal representations to steer model behavior. Its novel geometric decomposition framework offers deep mechanistic insights applicable across various LLMs and VLMs. While Paper 1 introduces a valuable benchmark for GUI agents, Paper 2 provides foundational knowledge that could broadly influence model interpretability, alignment, and future architecture design across multiple subfields of AI research.

Paper 2 likely has higher impact: it proposes a new autonomous RL training paradigm (co-evolving policy + training harness), targets a timely bottleneck in agentic RL (reward misalignment/hidden failure modes), and reports competitive or better results on high-value benchmarks including long-horizon software engineering—high real-world applicability and broad relevance to RL, agent design, and LLM training pipelines. Paper 1 is methodologically careful and insightful for mechanistic interpretability, but its applications are more indirect and its influence may be narrower than a training framework that can materially improve autonomous agents.

Paper 2 addresses a fundamental, domain-agnostic problem in AI interpretability by explaining the underlying mechanisms of how prompts steer foundation models. Its rigorous, causally tested geometric decomposition framework offers broad implications for understanding and improving LLMs and VLMs across various fields. In contrast, Paper 1 presents a highly specialized, domain-specific application for financial auditing, which, while valuable practically, has a narrower scope of scientific impact compared to the foundational insights provided by Paper 2.

Paper 2 addresses a fundamental 'black box' problem by introducing a novel geometric framework to mechanistically explain how prompting alters internal model representations. Its cross-modal applicability (LLMs and VLMs) and causal testing offer deep insights into model workings, promising broader long-term scientific impact across representational learning and AI interpretability compared to the specific, though important, behavioral safety vulnerability explored in Paper 1.

Paper 1 offers a novel, mechanistic framework for understanding how prompting alters internal representations in foundation models. This fundamental research in mechanistic interpretability has broad implications for AI alignment, capability steering, and general LLM architecture. In contrast, Paper 2 proposes a domain-specific benchmark for graph theory which, while useful for evaluation, provides less foundational innovation and has a narrower scope of impact.

Paper 2 likely has higher scientific impact due to its general, mechanistic contribution to understanding how prompting alters internal representations across multiple LLMs/VLMs, tasks, and modalities. The nested geometric decomposition plus causal layerwise state-mapping tests provide strong methodological rigor and a broadly applicable analysis toolkit for interpretability, controllability, and model design. Its relevance is high given widespread reliance on prompting. Paper 1 is impactful for drug discovery, but it is more domain-specific and its “agentic” component is less fundamentally novel than the cross-model causal geometry framework in Paper 2.

Paper 1 provides a foundational, causally-tested framework for understanding the internal mechanisms of prompting across both LLMs and VLMs. While Paper 2 offers a valuable method for scalable oversight, Paper 1's geometric decomposition addresses a fundamental black-box problem in AI interpretability. Its rigorous approach and broad applicability to multiple modalities and models offer deeper theoretical insights that could drive widespread innovations in model steering and design.

Paper 1 introduces a novel geometric decomposition framework for understanding how prompting reshapes internal representations in LLMs/VLMs, offering fundamental mechanistic insights with broad applicability across model architectures and tasks. Its rigorous methodology (nested geometric tiers with causal interventions) and interpretability contributions address a core question in AI alignment and interpretability research. Paper 2 addresses a practical but narrower concern—handoff costs in coding agents—with useful but incremental findings about efficiency gains from context-bearing handoffs. Paper 1's theoretical depth, methodological innovation, and breadth of impact across the rapidly growing interpretability field give it substantially higher scientific impact potential.

Paper 1 introduces a novel geometric decomposition framework that provides fundamental mechanistic insights into how prompting reshapes internal representations in LLMs and VLMs. Its contribution is broadly applicable across models, modalities, and tasks, offering interpretable causal analysis of prompt mechanisms. Paper 2 contributes a useful benchmark and incremental training improvements but is more narrowly focused on mathematical reasoning evaluation. Paper 1's framework has greater potential to influence interpretability, mechanistic understanding, and prompt engineering research across the field.

Paper 1 is more novel and broadly impactful: it introduces an interpretable, layered geometric decomposition with causal hidden-state mapping tests to explain how prompts transform internal representations across multiple LLMs/VLMs and diverse tasks. This offers a general mechanistic framework relevant to interpretability, controllability, alignment, and multimodal modeling. Paper 2 addresses an important, timely applied problem (tool overuse) with pragmatic RL refinements and solid benchmark gains, but its methodological novelty is more incremental and its impact is narrower (agent-tool policy optimization).

Paper 2 offers a novel, broadly applicable mechanistic framework for how prompting changes internal representations, with causal tests across multiple LLMs/VLMs and tasks. Its decomposition (from translation to nonlinear maps) yields interpretable, general insights (e.g., affine mixing as a key mechanism) that can influence prompting theory, representation learning, interpretability, and model steering—high breadth and timeliness. Paper 1 is important and timely for evaluation integrity, but is primarily a diagnostic/limitations study of existing contamination detectors; its impact is more specialized and less conceptually generative than Paper 2’s mechanistic, cross-domain methodology.

Paper 2 addresses a fundamental mechanism of how prompting reshapes internal representations in LLMs and VLMs. This provides foundational insights into AI behavior with broad applicability across all domains. While Paper 1 presents a rigorous and valuable domain-specific evaluation tool for chemistry, Paper 2's theoretical contributions to interpretability and model steering offer a significantly wider scientific impact.

Paper 2 addresses a fundamental scientific question about how prompting mathematically alters internal representations in LLMs and VLMs. This contribution to mechanistic interpretability provides foundational insights that apply broadly across AI research. In contrast, Paper 1 offers an innovative but more specialized engineering framework for self-improving agent systems.

Paper 2 introduces a novel, interpretable geometric decomposition framework for understanding how prompting steers LLM/VLM behavior—a fundamental question in AI interpretability. Its breadth (multiple models, modalities, and datasets) and mechanistic insights (identifying affine transformation as a key tier) have wide implications for understanding and improving foundation models. Paper 1 addresses an important but narrower robotics problem (parameterized diffusion policies). While valuable, Paper 2's contribution to mechanistic interpretability of LLMs/VLMs is more timely and broadly impactful across the field.

Paper 2 introduces a principled geometric decomposition framework for understanding how prompting steers LLM/VLM behavior internally, with broad applicability across multiple models, modalities, and tasks. It provides mechanistic interpretability insights (e.g., affine transformations as key mechanisms) with causal validation. Paper 1 addresses a narrower problem—tracking behavioral trajectories of agents via skill file diffs—with a simpler methodology (linear projection in embedding space) evaluated on a single trait with limited data (68 pairs). Paper 2's broader scope, deeper mechanistic insights, and relevance to fundamental LLM understanding give it substantially higher potential impact.

Paper 2 likely has higher impact: it proposes a concrete, interoperable protocol layer for autonomous science that could become infrastructure across many labs, instruments, and vendors. The primitives (capability signing, reservations, safety handshakes, physically typed/uncertainty-aware results) address real-world constraints (safety, exclusivity, metrology, reproducibility) and align with emerging standards (MCP/A2A), increasing adoption potential and timeliness. Paper 1 is methodologically rigorous and novel for mechanistic interpretability, but its applications are primarily analytical within ML, whereas LAP could reshape cross-disciplinary experimental automation.