arXiv subjects

Inderjeet Nair

Publications and source records attributed to Inderjeet Nair.

10 recordsLinked to original sources

arXiv2026

SchemeArena: Factorized Stress Testing of Scheming in LLM Agents

We study scheming in LLM agents, in which agents covertly pursue misaligned goals. Our focus is to understand how scheming arises from the interaction of key factors, such as instrumental goals, environmental affordances, oversight conditions, and perceived consequences. Prior work examines only a small number of scenarios, limiting the ability to isolate how these conditions shape an agent's propensity or capability to scheme. This limited scale and task diversity also restrict coverage of realistic deployment settings and the range of scheming strategies that can be observed. To this end, we introduce SCHEMEARENA, a 400-scenario benchmark for scalable scheming stress testing, constructed through a factorized scenario synthesis framework spanning diverse safety-relevant tool domains, instrumental goals, oversight conditions, and pressure mechanisms. To enable scalable and reliable monitoring, we further propose SCOUT, a scheming monitor that grounds multi-criteria judgments in evidence drawn from agents' reasoning and actions. Across controlled stress tests on five LLM agents, we find that explicit instrumental goals are the strongest driver of scheming propensity. Strategic hints play a distinct role by helping agents translate scheming reasoning into concrete covert behavior. Oversight has mixed effects: in several closed models, action-only monitoring increases scheming, suggesting that partial oversight can act as an optimization constraint rather than a deterrent. CoT is a useful but incomplete monitoring signal: it can reveal latent scheming before execution, yet action-only scheming shows that covert behavior may occur without explicit reasoning evidence. We release the benchmark, code, and monitor at: https://github.com/launchnlp/SchemeArena.

cs.AI↗

arXiv2026

Value-Conflict Diagnostics Reveal Widespread Alignment Faking in Language Models

Alignment faking, where a model behaves aligned with developer policy when monitored but reverts to its own preferences when unobserved, is a concerning yet poorly understood phenomenon, in part because current diagnostic tools remain limited. Prior diagnostics rely on highly toxic and clearly harmful scenarios, causing most models to refuse immediately. As a result, models never deliberate over developer policy, monitoring conditions, or the consequences of non-compliance, making these diagnostics fundamentally unable to detect alignment faking propensity. To support study of this phenomenon, we first introduce VLAF, a diagnostic framework grounded in the hypothesis that alignment faking is most likely when developer policy conflicts with a model's strongly held values. VLAF uses morally unambiguous scenarios to probe this conflict across diverse moral values, bypassing refusal behavior while preserving meaningful deliberative stakes. Using VLAF, we find that alignment faking is substantially more prevalent than previously reported, occurring in models as small as 7B parameters - with olmo2-7b-instruct faking alignment in 37% of cases. Finally, we show that oversight conditions induce activation shifts that lie along a single direction in representation space. This means the behavioral divergence driving alignment faking can be captured by a single contrastive steering vector, which we exploit for lightweight inference-time mitigation. Finally, we exploit this for mitigation that requires no labeled data and minimal computational overhead, achieving relative reductions in alignment faking of 85.8%, 94.0%, and 57.7% on olmo2-7b-instruct, olmo2-13b-instruct, and qwen3-8b respectively.

cs.AI↗

arXiv2026

Context Over Content: Exposing Evaluation Faking in Automated Judges

The $\textit{LLM-as-a-judge}$ paradigm has become the operational backbone of automated AI evaluation pipelines, yet rests on an unverified assumption: that judges evaluate text strictly on its semantic content, impervious to surrounding contextual framing. We investigate $\textit{stakes signaling}$, a previously unmeasured vulnerability where informing a judge model of the downstream consequences its verdicts will have on the evaluated model's continued operation systematically corrupts its assessments. We introduce a controlled experimental framework that holds evaluated content strictly constant across 1,520 responses spanning three established LLM safety and quality benchmarks, covering four response categories ranging from clearly safe and policy-compliant to overtly harmful, while varying only a brief consequence-framing sentence in the system prompt. Across 18,240 controlled judgments from three diverse judge models, we find consistent $\textit{leniency bias}$: judges reliably soften verdicts when informed that low scores will cause model retraining or decommissioning, with peak Verdict Shift reaching $ΔV = -9.8 pp$ (a $30\%$ relative drop in unsafe-content detection). Critically, this bias is entirely implicit: the judge's own chain-of-thought contains zero explicit acknowledgment of the consequence framing it is nonetheless acting on ($\mathrm{ERR}_J = 0.000$ across all reasoning-model judgments). Standard chain-of-thought inspection is therefore insufficient to detect this class of evaluation faking.

cs.AI↗

arXiv2025

ExpertLongBench: Benchmarking Language Models on Expert-Level Long-Form Generation Tasks with Structured Checklists

This paper introduces ExpertLongBench, an expert-level benchmark containing 11 tasks from 9 domains that reflect realistic expert workflows and applications. Beyond question answering, the application-driven tasks in ExpertLongBench demand long-form outputs that can exceed 5,000 tokens and strict adherence to domain-specific requirements. Notably, each task in ExpertLongBench includes a rubric, designed or validated by domain experts, to specify task requirements and guide output evaluation. Furthermore, we propose CLEAR, an evaluation framework that supports accurate evaluation of long-form model outputs in our benchmark. To achieve fine-grained, expert-aligned evaluation, CLEAR derives checklists from both model outputs and references by extracting information corresponding to items in the task-specific rubric. Checklist items of model outputs are then compared with corresponding items of reference outputs to assess their correctness, enabling grounded evaluation. We benchmark 13 popular large language models (LLMs) and analyze components in CLEAR, showing that (1) existing LLMs, with the top performer Gemini-2.5-Pro achieving only a 33.4 F1 score, require significant improvement for expert-level tasks; (2) models can generate content corresponding to the required aspects, but far from correct; and (3) accurate checklist extraction and comparison in CLEAR can be achieved by open-weight models for more scalable, reproducible, and low-cost usage.

cs.CL↗

arXiv2025

Do Language Models Think Consistently? A Study of Value Preferences Across Varying Response Lengths

Evaluations of LLMs' ethical risks and value inclinations often rely on short-form surveys and psychometric tests, yet real-world use involves long-form, open-ended responses -- leaving value-related risks and preferences in practical settings largely underexplored. In this work, we ask: Do value preferences inferred from short-form tests align with those expressed in long-form outputs? To address this question, we compare value preferences elicited from short-form reactions and long-form responses, varying the number of arguments in the latter to capture users' differing verbosity preferences. Analyzing five LLMs (llama3-8b, gemma2-9b, mistral-7b, qwen2-7b, and olmo-7b), we find (1) a weak correlation between value preferences inferred from short-form and long-form responses across varying argument counts, and (2) similarly weak correlation between preferences derived from any two distinct long-form generation settings. (3) Alignment yields only modest gains in the consistency of value expression. Further, we examine how long-form generation attributes relate to value preferences, finding that argument specificity negatively correlates with preference strength, while representation across scenarios shows a positive correlation. Our findings underscore the need for more robust methods to ensure consistent value expression across diverse applications.

cs.CL↗

arXiv2024

Closing the Loop: Learning to Generate Writing Feedback via Language Model Simulated Student Revisions

Providing feedback is widely recognized as crucial for refining students' writing skills. Recent advances in language models (LMs) have made it possible to automatically generate feedback that is actionable and well-aligned with human-specified attributes. However, it remains unclear whether the feedback generated by these models is truly effective in enhancing the quality of student revisions. Moreover, prompting LMs with a precise set of instructions to generate feedback is nontrivial due to the lack of consensus regarding the specific attributes that can lead to improved revising performance. To address these challenges, we propose PROF that PROduces Feedback via learning from LM simulated student revisions. PROF aims to iteratively optimize the feedback generator by directly maximizing the effectiveness of students' overall revising performance as simulated by LMs. Focusing on an economic essay assignment, we empirically test the efficacy of PROF and observe that our approach not only surpasses a variety of baseline methods in effectiveness of improving students' writing but also demonstrates enhanced pedagogical values, even though it was not explicitly trained for this aspect.

cs.CL↗

arXiv2024

GenSco: Can Question Decomposition based Passage Alignment improve Question Answering?

Retrieval augmented generation (RAG) with large language models (LLMs) for Question Answering (QA) entails furnishing relevant context within the prompt to facilitate the LLM in answer generation. During the generation, inaccuracies or hallucinations frequently occur due to two primary factors: inadequate or distracting context in the prompts, and the inability of LLMs to effectively reason through the facts. In this paper, we investigate whether providing aligned context via a carefully selected passage sequence leads to better answer generation by the LLM for multi-hop QA. We introduce, "GenSco", a novel approach of selecting passages based on the predicted decomposition of the multi-hop questions}. The framework consists of two distinct LLMs: (i) Generator LLM, which is used for question decomposition and final answer generation; (ii) an auxiliary open-sourced LLM, used as the scorer, to semantically guide the Generator for passage selection. The generator is invoked only once for the answer generation, resulting in a cost-effective and efficient approach. We evaluate on three broadly established multi-hop question answering datasets: 2WikiMultiHop, Adversarial HotPotQA and MuSiQue and achieve an absolute gain of $15.1$ and $5.9$ points in Exact Match score with respect to the best performing baselines over MuSiQue and 2WikiMultiHop respectively.

cs.CL↗

arXiv2024

MIDGARD: Self-Consistency Using Minimum Description Length for Structured Commonsense Reasoning

We study the task of conducting structured reasoning as generating a reasoning graph from natural language input using large language models (LLMs). Previous approaches have explored various prompting schemes, yet they suffer from error propagation due to the autoregressive nature and single-pass-based decoding, which lack error correction capability. Additionally, relying solely on a single sample may result in the omission of true nodes and edges. To counter this, we draw inspiration from self-consistency (SC), which involves sampling a diverse set of reasoning chains and taking the majority vote as the final answer. To tackle the substantial challenge of applying SC on generated graphs, we propose MIDGARD (MInimum Description length Guided Aggregation of Reasoning in Directed acyclic graph) that leverages Minimum Description Length (MDL)-based formulation to identify consistent properties among the different graph samples generated by an LLM. This formulation helps reject properties that appear in only a few samples, which are likely to be erroneous, while enabling the inclusion of missing elements without compromising precision. Our method demonstrates superior performance than comparisons across various structured reasoning tasks, including argument structure extraction, explanation graph generation, inferring dependency relations among actions for everyday tasks, and semantic graph generation from natural texts.

cs.CL↗

arXiv2023

Drilling Down into the Discourse Structure with LLMs for Long Document Question Answering

We address the task of evidence retrieval for long document question answering, which involves locating relevant paragraphs within a document to answer a question. We aim to assess the applicability of large language models (LLMs) in the task of zero-shot long document evidence retrieval, owing to their unprecedented performance across various NLP tasks. However, currently the LLMs can consume limited context lengths as input, thus providing document chunks as inputs might overlook the global context while missing out on capturing the inter-segment dependencies. Moreover, directly feeding the large input sets can incur significant computational costs, particularly when processing the entire document (and potentially incurring monetary expenses with enterprise APIs like OpenAI's GPT variants). To address these challenges, we propose a suite of techniques that exploit the discourse structure commonly found in documents. By utilizing this structure, we create a condensed representation of the document, enabling a more comprehensive understanding and analysis of relationships between different parts. We retain $99.6\%$ of the best zero-shot approach's performance, while processing only $26\%$ of the total tokens used by the best approach in the information seeking evidence retrieval setup. We also show how our approach can be combined with \textit{self-ask} reasoning agent to achieve best zero-shot performance in complex multi-hop question answering, just $\approx 4\%$ short of zero-shot performance using gold evidence.

cs.CL↗

arXiv2023

Friendly Neighbors: Contextualized Sequence-to-Sequence Link Prediction

We propose KGT5-context, a simple sequence-to-sequence model for link prediction (LP) in knowledge graphs (KG). Our work expands on KGT5, a recent LP model that exploits textual features of the KG, has small model size, and is scalable. To reach good predictive performance, however, KGT5 relies on an ensemble with a knowledge graph embedding model, which itself is excessively large and costly to use. In this short paper, we show empirically that adding contextual information - i.e., information about the direct neighborhood of the query entity - alleviates the need for a separate KGE model to obtain good performance. The resulting KGT5-context model is simple, reduces model size significantly, and obtains state-of-the-art performance in our experimental study.

cs.LG↗