{"id":"b3ec68f9-f407-45ce-9b0e-8bbd82f21cc5","arxiv_id":"2605.29561","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"ParaTool pre-trains independent parameter modules for tools, uses a gating network for dynamic selection, and fine-tunes them so LLMs can call tools without in-context documentation.","lead":"ParaTool represents each tool as a dedicated set of parameters that an LLM can load dynamically instead of reading tool docs in its prompt. If it works, this could cut inference cost and hallucination risk for tool-using AI systems.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Independent tool parameter modules may fail to capture query-dependent or inter-tool usage patterns that in-context examples provide directly.","rationale":"The reader's weakest assumption directly targets the encapsulation claim. The concern above refines it to the specific case of context- or composition-dependent usage, which is a concrete risk for the 'without in-context documents' claim and is testable on the reported benchmarks without requiring new experiments.","tokens_in":1708,"tokens_out":270,"duration_ms":26723,"concrete_test":"On the BFCL subset requiring multi-tool composition, compare ParaTool accuracy against the ICL baseline when the same tools appear both singly and in combination; if the relative gain disappears or reverses on the composition subset, the independence assumption does not hold.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires that pre-training can fully encapsulate each tool's usage into standalone, loadable parameter sets, with a gating network then selecting and combining them at inference. This implicitly assumes tool behaviors are modular and context-independent enough for soft aggregation to substitute for explicit documentation. If tool invocation rules depend on the current query state, prior tool outputs, or combinations (common in BFCL multi-tool tasks), the independent modules plus gating may not reconstruct the necessary conditional logic without additional cross-module mechanisms.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes ParaTool, a three-stage framework for LLM tool calling that (1) pre-trains independent parameter modules to encapsulate each tool's knowledge, (2) uses a gating network for soft selection and aggregation of relevant modules at inference, and (3) jointly fine-tunes the modules. The central claim is that this shifts tool representations from in-context documents/examples to parameters, enabling tool calling without ICL overhead while outperforming strong ICL baselines on Stable ToolBench and BFCL with lower computational complexity.","tokens_in":1808,"tokens_out":472,"duration_ms":22003,"significance":"If the performance claims are substantiated with detailed results, the approach could meaningfully reduce context-length costs and hallucination risks in tool-augmented LLMs by making tool knowledge modular and loadable. The gating-based aggregation is a potentially scalable alternative to explicit documentation, though its success hinges on the modularity assumption.","major_comments":[{"comment":"Abstract and Experiments section: the central claim of 'significantly outperforms strong ICL-based baselines' on Stable ToolBench and BFCL is unsupported because no quantitative results, error bars, ablation studies, or descriptions of baseline implementations (e.g., how context length or example selection was controlled) are provided, making the outperformance and complexity-reduction assertions unevaluable.","section":"Abstract / Experiments"},{"comment":"Method (parametric tool pre-training and soft tool selection stages): the framework assumes tool behaviors can be fully captured in independent, query-independent parameter modules whose soft aggregation via gating substitutes for in-context examples; this is load-bearing for BFCL multi-tool tasks, yet no cross-module interaction mechanisms or tests for query-dependent / inter-tool conditional logic are described, leaving open whether the modules can reconstruct necessary state-dependent invocation rules.","section":"Method"}],"minor_comments":[{"comment":"The three-stage pipeline description would benefit from an explicit diagram or pseudocode showing how the gating network interfaces with the base LLM forward pass.","section":null}],"recommendation":"major_revision","confidential_remarks":"The manuscript's experimental section appears to omit the quantitative evidence required to assess its primary claims; this is a basic presentation issue that should be addressed before further review."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback. The comments identify key gaps in the presentation of results and methodological details that we will address through revision.","responses":[{"response":"We agree that the submitted manuscript does not provide the quantitative results, error bars, ablation studies, or baseline implementation details needed to evaluate the claims. In the revised manuscript we will add these elements to the experiments section (including tables with metrics and standard deviations, ablation results, and explicit descriptions of ICL baseline setups such as context length and example selection) and update the abstract to reference key findings.","revision_made":"yes","referee_comment":"[Abstract / Experiments] Abstract and Experiments section: the central claim of 'significantly outperforms strong ICL-based baselines' on Stable ToolBench and BFCL is unsupported because no quantitative results, error bars, ablation studies, or descriptions of baseline implementations (e.g., how context length or example selection was controlled) are provided, making the outperformance and complexity-reduction assertions unevaluable."},{"response":"The gating network is intended to enable soft aggregation of multiple modules for multi-tool scenarios. We acknowledge that explicit cross-module interaction mechanisms and dedicated tests for query-dependent conditional logic are not described. In revision we will expand the method section with additional discussion of how joint fine-tuning supports inter-tool coordination and include relevant analysis or results from BFCL multi-tool cases; we will also note this as an area for future extension if needed.","revision_made":"partial","referee_comment":"[Method] Method (parametric tool pre-training and soft tool selection stages): the framework assumes tool behaviors can be fully captured in independent, query-independent parameter modules whose soft aggregation via gating substitutes for in-context examples; this is load-bearing for BFCL multi-tool tasks, yet no cross-module interaction mechanisms or tests for query-dependent / inter-tool conditional logic are described, leaving open whether the modules can reconstruct necessary state-dependent invocation rules."}],"tokens_in":1391,"tokens_out":423,"duration_ms":24259,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main takeaway is that ParaTool shifts tool knowledge into parameter modules via pre-training, gating, and fine-tuning, but the abstract provides no data to back up the claimed gains over in-context methods.\n\nThe three-stage setup is the clearest new piece: first encapsulate each tool into its own parameters, then train a gating network to weigh and combine them on the fly, then jointly fine-tune so training and inference line up. This is more explicit about modularity than generic adapter or prompt-tuning work, and it directly attacks the context-length problem that grows with tool docs and examples.\n\nThe paper does identify a practical issue—ICL for tool calling inflates inference cost and hallucination risk—and the parameter route is a logical alternative if the modules can carry the necessary usage patterns.\n\nThe soft spot is the missing evidence. The abstract says ParaTool beats strong ICL baselines on Stable ToolBench and BFCL with lower complexity, yet gives no scores, no error bars, no baseline implementation details, and no ablation results. That makes the central performance claim impossible to assess from what is here.\n\nA second concern is whether independent modules plus soft gating can handle query-dependent or multi-tool logic. BFCL tasks often require conditional behavior based on prior outputs or the current state; if the modules stay fully separate, the gating may not reconstruct those interactions without extra cross-module machinery. The stress-test note flags this, and nothing in the abstract rules it out.\n\nThis is for researchers focused on efficient tool-augmented agents who want to test whether parameterizing tools can replace context stuffing. A reader looking for concrete efficiency wins would need the full experimental section.\n\nI would send it for peer review so the numbers, ablations, and any handling of inter-tool dependencies can be checked.","headline":"ParaTool's three-stage shift of tool knowledge into loadable parameter modules targets ICL overhead but the abstract supplies no numbers or ablations to show whether the gains are real.","tokens_in":2284,"tokens_out":444,"would_cite":false,"duration_ms":20612,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"LLMs can call tools using dedicated loadable parameter modules instead of in-context documentation or examples.","keywords":["tool calling","parameterized tools","large language models","in-context learning","gating network","tool pre-training","fine-tuning"],"falsifier":"An experiment in which ParaTool's accuracy falls below in-context baselines when the set of tools grows or when tool instructions are long and nuanced would falsify the claim.","tokens_in":2617,"feed_emoji":"🛠","tokens_out":565,"duration_ms":24545,"temperature":0.7,"pith_summary":"Mainstream tool-calling methods either embed full tool documentation and examples in the prompt, which raises inference costs and hallucination risk, or fine-tune the model yet still require context for tool specifics. ParaTool instead encodes each tool as its own independent parameter module during a dedicated pre-training stage. A gating network then selects and combines the relevant modules at inference time. A final joint fine-tuning stage aligns the process. On Stable ToolBench and BFCL the method outperforms strong in-context baselines while lowering computational load.","feed_headline":"Tool knowledge moves from prompts into loadable parameters","feed_subtitle":"ParaTool encodes tools as separate parameter modules, gates them at inference, and beats ICL baselines on benchmarks with lower compute.","key_machinery":"Dedicated loadable parameter modules for each tool, dynamically selected and aggregated by a gating network.","core_discovery":"ParaTool projects each tool into a dedicated, loadable set of parameters. Parametric tool pre-training encapsulates tool knowledge into independent modules; soft tool selection uses a gating network to weigh and aggregate relevant parameters; and parametric tool fine-tuning jointly updates the modules to align training and inference. This lets the LLM perform tool calling without relying on in-context documents or examples.","pith_inferences":["New tools could be added by training only their parameter modules without retraining the base model.","The same parameterization approach might apply to other external knowledge sources such as APIs or databases.","Inference-time scaling to thousands of tools becomes possible if only relevant modules are activated."],"forward_implications":["Tool calling incurs lower inference cost because tool documentation no longer expands the prompt.","Hallucination risk tied to long contexts is reduced.","Performance on Stable ToolBench and BFCL exceeds that of strong ICL baselines.","Tool modules can be pre-trained independently and loaded only when needed."],"fun_headline_variants":["ParaTool stores tools as loadable parameter modules","Tools move from context into dedicated parameter sets","Gating network selects relevant tool parameters dynamically","Parametric pre-training replaces in-context tool examples"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Tool-specific knowledge can be fully captured inside separate parameter modules that a gating network can select and combine at inference time.","fun_headline_variants_meta":{"raw":{"variants":["ParaTool stores tools as loadable parameter modules","Tools move from context into dedicated parameter sets","Gating network selects relevant tool parameters dynamically","Parametric pre-training replaces in-context tool examples"]},"model":"grok-4.3","cost_usd":0.004052,"raw_usage":{"total_tokens":2062,"prompt_tokens":667,"num_sources_used":0,"completion_tokens":55,"cost_in_usd_ticks":40524500,"prompt_tokens_details":{"text_tokens":667,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1340,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":667,"tokens_out":55,"duration_ms":10302,"temperature":1.0,"reasoning_tokens":1340,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T07:54:34.393328+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"An experiment in which ParaTool's accuracy falls below in-context baselines when the set of tools grows or when tool instructions are long and nuanced would falsify the claim.","supporting_citations":[],"review_version":1}