Beyond Fluency: Toward Reliable Trajectories in Agentic IR DOI: 10.1145/nnnnnnn.nnnnnnn Conference: SIGIR 2026; July 2026; Gold Coast, Australia Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval Price: 15.00 ISBN: 978-1-4503-XXXX-X/26/07 CCS: Information systems Users and interactive retrieval CCS: Computing methodologies Natural language processing CCS: Computing methodologies Planning and scheduling CCS: Security and privacy Software security engineering
Anushree Sinha, Srivaths Ranganathan, Debanshu Das, Abhishek Dharmaratnakar · 2026
Read Beyond Fluency: Toward Reliable Trajectories in Agentic IR DOI: 10.1145/nnnnnnn.nnnnnnn Conference: SIGIR 2026; July 2026; Gold Coast, Australia Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval Price: 15.00 ISBN: 978-1-4503-XXXX-X/26/07 CCS: Information systems Users and interactive retrieval CCS: Computing methodologies Natural language processing CCS: Computing methodologies Planning and scheduling CCS: Security and privacy Software security engineering by Anushree Sinha, Srivaths Ranganathan, Debanshu Das, Abhishek Dharmaratnakar (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-04269.txt.
Introduction
Sinha and colleagues track a shift in search tech, from ranking documents toward agents that reason, act and observe over many steps. In such long runs, small early mistakes can snowball, leaving fluent language that masks a mismatch between thought and tool use. This position paper sorts failures seen in work systems across planning, retrieval, reasoning and execution. It calls for checks at each step and holding back under doubt, putting correct process ahead of smooth but unverified completion.
Catalog record
| CREATOR | Anushree Sinha, Srivaths Ranganathan, Debanshu Das, Abhishek Dharmaratnakar |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 20186 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.04269
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Beyond Fluency: Toward Reliable Trajectories in Agentic IR DOI: 10.1145/nnnnnnn.nnnnnnn Conference: SIGIR 2026; July 2026; Gold Coast, Australia Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval Price: 15.00 ISBN: 978-1-4503-XXXX-X/26/07 CCS: Information systems Users and interactive retrieval CCS: Computing methodologies Natural language processing CCS: Computing methodologies Planning and scheduling CCS: Security and privacy Software security engineering Anushree Sinha, Srivaths Ranganathan, Debanshu Das, Abhishek Dharmaratnakar Abstract . Information Retrieval is shifting from passive document ranking toward autonomous agentic workflows that operate in multi-step Reason–Act–Observe loops ( 30 ) . In such long-horizon trajectories, minor early errors can cascade, leading to functional misalignment between internal reasoning and external tool execution despite continued linguistic fluency. This position paper synthesizes failure modes observed in industrial agentic systems, categorizing errors across planning, retrieval, reasoning, and execution. We argue that safe deployment requires moving beyond endpoint accuracy toward trajectory integrity and causal attribution. To address compounding error and deceptive fluency, we propose verification gates at each interaction unit and advocate systematic abstention under calibrated uncertainty. Reliable Agentic IR systems must prioritize process correctness and grounded execution over plausible but unverified completion. 1. Introduction: From Retrieval to Action Information Retrieval (IR) is undergoing a fundamental shift. We are moving away from passive systems that return a list of ranked documents toward autonomous agentic workflows that interleave internal deliberation with external actions ( 49 ; 5 ) . These systems, often referred to as Agentic AI, operate in multi-step Reason-Act-Observe loops, calling APIs, executing code, and interacting with dynamic environments to fulfill complex user intents ( 46 ; 7 ; 30 ; 6 ) . This transition extends classical models of iterative and session-based information seeking, where user goals evolve over time rather than being specified in a single static query ( 1 ; 16 ; 31 ) . In a standard document-retrieval setting, success is measured by the relevance of a static result set. In an agentic loop, however, success depends on the trajectory reliability of information use over time. If an agent retrieves a perfectly relevant document at the start of a task but misinterprets a technical constraint during a subsequent API call ( 29 ) , the resulting trajectory collapse invalidates the initial retrieval success. The primary bottleneck for the industrial adoption of these agents is the phenomenon of error propagation, often described as the Snowball Effect ( 24 ; 47 ) . Unlike hallucinations in static tasks ( 17 ) , agentic hallucinations are dynamic; they emerge when a minor logical divergence in an initial search or planning step cascades through subsequent tool calls ( 24 ; 41 ) . Our position emphasizes that the industry is currently caught in a fluency trap, where the high linguistic coherence of Large Language Models (LLMs) ( 4 ; 37 ) masks deep-seated functional misalignments. An agent may produce a grammatically perfect and seemingly logical "Chain of Thought" while simultaneously fabricating non-existent API arguments ( 28 ) or ignoring environment feedback ( 35 ) . This paper provides a comprehensive synthesis of these failure modes. For autonomous agents to be safely deployed in production, we must prioritize automated verification gates that ensure every step of a reasoning chain is factually grounded. Table 1. Comprehensive Process-Oriented Taxonomy of Agentic Hallucinations ( 29 ) . Functional Stage Hallucination Category Sub-category Operational Description Planning Structural Failure Fact Derive The agent introduces non-existent or misleading facts during initial strategy formation. Task Decompose The agent produces task-misaligned subgoals that ignore primary constraints ( 14 ) . Retrieval Contextual Failure Query Misalign The agent formulates inaccurate retrieval queries that miss relevant information ( 22 ) . Context Misalign The agent retrieves factually incorrect context as truth, overriding parametric knowledge ( 39 ) . Summarize Misalign The agent misrepresents retrieved documents via inaccurate internal summarization ( 52 ) . Reasoning Internal Logic Failure Factual Reasoning The agent makes incorrect logical inferences over provided context ( 46 ) . Math Reasoning The agent performs incorrect calculations or algebraic derivations ( 42 ) . Science Reasoning The agent generates incorrect scientific inferences or symbolic interpretations. Solvability Hallucination The agent misjudges task complexity, believing an unsolvable task is solvable ( 51 ; 15 ) . Action Execution Failure Missing Tool The agent fails to invoke a tool when one is strictly necessary for success. Incorrect Argument The agent specifies wrong parameters/arguments for a valid tool invocation ( 29 ) . Benchmark environments that expose agents to realistic action spaces—such as WebShop, WebArena, Mind2Web, and SWE-bench—demonstrate how minor parameter errors can cascade into complete task failure in practical settings ( 45 ; 54 ; 9 ; 18 ) . Parallel Conflict The agent triggers execution errors via conflicting or redundant parallel actions ( 53 ; 30 ) . Unnecessary Tool The agent invokes irrelevant tools that waste budget or risk side effects ( 32 ) . 2. Mapping the Failure Surface To address the reliability gap in production, we must categorize failures not by their linguistic form, but by their point of entry within the agentic lifecycle. We categorize the existing literature into four functional stages where logical divergences typically originate. 2.1. Planning and Intent Decomposition The first point of failure occurs during the high-level cognitive phase where an agent decomposes a user’s global intent ( ( 1 ) into executable sub-goals. Industrial workflows often break when an agent introduces misleading premises during initial strategy formation. Research has demonstrated that LLMs frequently struggle with task decomposition, often failing to solve even relatively simple logical puzzles when they require multi-step lookahead ( 19 ) . These planning gaps are further complicated by "solvability hallucinations," where agents attempt to force a plan for a query that is inherently impossible to complete with the provided toolset ( 51 ) . This initial misalignment in goal parameters sets the foundation for the cascading errors that follow ( 14 ) . Step 1: Planning (Fact Derive Error) Misdefines goal parameters. Step 2: Action (Tool Execution) Obs Invokes tool with erroneous parameters. Step 3: Thought (Reasoning) Obs Causal divergence accumulates. Final Result (Hallucination) Trajectory failure despite fluency. Figure 1. The Snowball Effect in Agentic Trajectories: Visualizing causal error propagation . 2.2. Retrieval and Contextual Integration In Retrieval-Augmented Generation (RAG) agents, the interface between the model’s parametric memory and external knowledge bases is a primary site for "contextual drift." Similar challenges have been observed in conversational search settings, where multi-turn context accumulation and query reformulation introduce additional opportunities for misalignment ( 8 ) . Even when relevant information is retrieved, agents often suffer from "prior-bias," where internal model weights override the provided evidence, leading to incorrect synthesis ( 22 ) . Furthermore, in dynamic environments, the temporal decay of information becomes a critical factor; agents often fail to refresh their internal state when search engine results contradict their training data ( 39 ) . This failure to properly integrate retrieved context often manifests as "Summarize Misalign," where the agent’s internal synthesis misrepresents the very documents it successfully retrieved ( 17 ; 10 ) . 2.3. Internal Reasoning Hallucinations A distinct class of failure occurs during the deliberation phase, where the agent processes retrieved information to form a logical