Papers for

llm application developers

Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.

Large language model choices shift when extra context is added

Reading Too Much into Context: Passive Exposure Can Steer LLM Decisions

Abstract: Large language model (LLM) assistants can now search the web and consult external sources while completing user requests. These sources can provide useful evidence, but they can also introduce additional content into the model's context. Can such passive exposure steer a decision even when the added content provides no reason to change it? We examine the stability of model decisions on the same tasks with and without such external content. Across all open-weight and closed-weight models we test, exposure systematically shifts decisions, with effects reaching nearly 50 percentage points in closed-weight models. The same pattern appears with real-world online opinions. The influence also extends beyond subjective preferences. Such exposure can steer models toward choices that violate explicit user requirements and increase their acceptance of false claims. In short, what enters an LLM's context can influence its decision even when it should not determine it.

Sun 27 SeptComputation and LanguageMachine Learning
The gist
Large language models (LLMs) can look up information from outside sources while answering questions. The authors found that simply adding extra content to the model’s context can change its decisions, even if that content shouldn’t affect the answer. This change in behavior happens across different models and can cause the model to ignore user instructions or believe false information. The study shows that what the model sees matters a lot, sometimes too much, when making decisions.
Open → 2609.33065v1

Authorization closure graph improves updates for llm agents

Authorization Closure Graph: Minimal Repair for LLM Agents with Evolving User Instructions

Abstract: Tool-using large language model (LLM) agents increasingly perform state-changing actions that require user authorization. Yet existing approaches do not provide a principled mechanism for selectively updating prior authorization when only part of an instruction changes. To this end, we propose an Authorization-Closure-Graph (ACG)-based framework that represents authorization and its dependencies as an evolving, versioned state. ACG selectively invalidates authority affected by a revision while preserving unaffected portions of the authorization state, and computes a minimal repair that identifies only the missing evidence or authority required for execution. This enables agents to adapt to revised instructions while avoiding stale authority and unnecessary authorization requests. We evaluate ACG across three advanced LLMs in two natural tasks, and ACG consistently improves action safety rate and task success rate. Code is available at https://github.com/weiliang822/ACG.

Sat 26 SeptArtificial Intelligence
The gist
Large language model agents often need permission to make changes based on user instructions, but when users change just part of their request, existing methods don’t handle updating these permissions well. The authors propose a system, called Authorization-Closure-Graph (ACG), which tracks permissions and their connections as the instructions change. ACG smartly updates only the parts of permissions affected by revisions and figures out what extra permissions are needed, avoiding unnecessary or outdated authorization requests. Testing with different large language models on real tasks showed that ACG helps agents act more safely and successfully.
Open → 2609.32428v1