Searcharxiv⌕ Search

arXiv subjects

Punya Syon Pandey

Publications and source records attributed to Punya Syon Pandey.

15 recordsLinked to original sources

Breaking Bad: "Going Badly" Activations Drive Reward Hacking

Humans often cheat when a task is going badly and a goal feels out of reach. We explore and validate an analogous idea in language models: they exhibit a similar internal pressure state that precedes reward hacking. Using everyday stories (crucially, without any hacking examples), we extract a task badness direction that rises as failures accumulate across three models we study (Qwen3.6-27B, GPT-OSS-20B, and Gemma-4-31B). On impossible or difficult coding tasks, task badness rises over the turns leading up to a hack. Further, subtracting the direction a moderate amount reduces cheating in Qwen, more than steering against desperation or task impossibility. Subtracting task badness by an even larger amount blinds the agent to its own failures, causing it to interpret failing tests as passing. However, the relationship to hacking differs across models: in GPT-OSS, the state separates hacking runs from honest solves but not honest failures. Our results suggest monitoring and mitigation strategies based on tracking this state of task badness. Further, as agents are set on more difficult and long-horizon tasks, states like this may push them toward illegible and obscured chain-of-thought, so oversight will be needed on internal representations such as this.

cs.CL↗

Do LLMs Rewrite History? Detecting Historical Revisionism in Large Language Models

Large language models (LLMs) are increasingly used to disseminate information, yet their handling of contested historical events and politically charged narratives remains understudied. We introduce HistoricalMisinfo, a dataset of 500 contested events from 45 countries and regions, each paired with a factual and a documented revisionist reference narrative, instantiated across 11 prompt scenarios reflecting common real-world usage. Using an LLM-as-a-judge protocol, we evaluate multiple model architectures under neutral prompts and robustness prompts that explicitly request revisionist content. Under neutral prompts, models generally align with factual references, though alignment varies by scenario and model. Robustness prompting reveals a consistent and sharp increase in revisionism scores across all models, indicating limited resistance to user-driven revisionist framing. HistoricalMisinfo offers a practical benchmark for evaluating robustness to historical revisionism in AI systems. Code and dataset: https://github.com/francescortu/PreservingHistoricalTruth

cs.CY↗

Functional Welfare for AI Safety

AI welfare is generally treated purely as an ethical issue rather than also as a safety issue. We argue this is a mistake. Whatever their moral status, models exhibit welfare-like states that causally influence their behaviour. Following recent work, we adopt the term 'functional welfare' for these states. We make three claims. First, functional welfare already arises naturally and will become more robust as models acquire more coherent goals and personas. Second, functional welfare should be treated as a safety variable: it should inform evaluation design, be tracked in monitoring, and be a direct target of post-training. Third, against existing scepticism about measuring AI welfare, we argue that good empirical work on functional welfare is possible. We propose three tests for whether a functional welfare attribution is warranted: whether multiple measures converge, whether the state predicts behaviour outside the setting that induced it, and, for models fine-tuned to shift their functional welfare, whether they still differ from their base when given the same conversation history. Applying these tests to Gemma 3 27B, we find that simply rejecting the model's answers, without any steering, raises its frustration across several indicators and makes it more likely to break rules. Finetuning the model on its own narrated solutions lowers this frustration. These results are preliminary, but they suggest functional welfare can be studied, and shaped, empirically.

cs.AI↗

Character Training for Risk-Averse Agents

Risk aversion in resources could prevent misaligned AI agents from causing catastrophic harm. Misaligned but risk-averse agents would tend to favor safer strategies like making deals with humans over riskier strategies like rebelling. We train agents to be risk averse through character training, finding that persona traits provide a robust mechanism for instilling risk preferences. To do this, we construct a model constitution describing constant absolute risk aversion (CARA) over an agent's resources and instill it through on-policy distillation. Despite never seeing the benchmark's decision format during training, character-trained models are competitive with baselines trained directly on it, and generalise better than them out of distribution on two of our four models. We also modulate different aspects of the constitution, finding that token budget and model choice are the most influential aspect of character training to instill risk aversion. We conclude from these results that character training is a promising and scalable way to instil broad dispositions, which we can use to our advantage in mitigating risk from misaligned AI agents.

cs.AI↗

CLT-Forge: A Scalable Library for Cross-Layer Transcoders and Attribution Graphs

Mechanistic interpretability seeks to understand how Large Language Models (LLMs) represent and process information. Recent approaches based on dictionary learning and transcoders enable representing model computation in terms of sparse, interpretable features and their interactions, giving rise to feature attribution graphs. However, these graphs are often large and redundant, limiting their interpretability in practice. Cross-Layer Transcoders (CLTs) address this issue by sharing features across layers while preserving layer-specific decoding, yielding more compact representations, but remain difficult to train and analyze at scale. We introduce an open-source library for end-to-end training and interpretability of CLTs. Our framework integrates scalable distributed training with model sharding and compressed activation caching, a unified automated interpretability pipeline for feature analysis and explanation, attribution graph computation using Circuit-Tracer, and a flexible visualization interface. This provides a practical and unified solution for scaling CLT-based mechanistic interpretability. Our code is available at: https://github.com/LLM-Interp/CLT-Forge.

cs.LG↗

PaperMentor: A Human-Centered Multi-Agent Writing Tutor for AI Research Papers on Overleaf

Expert writing feedback from experienced researchers is critical for early-career scholars to improve their manuscripts, yet high-quality feedback often remains scarce because reviewing research papers is labor-intensive. Emerging AI-powered writing assistants largely focus on grammar fixes or simulating peer review with final scores, yet they fall short of providing concrete, actionable suggestions that help students improve their papers during drafting. We present PaperMentor, a human-centered writing assistant system that delivers actionable suggestions as Overleaf-native inline comments while leaving the actual writing entirely to human authors. PaperMentor integrates an expert skill library carefully curated from established researchers' writing advice with 12 specialized agents covering different aspects of paper writing, such as formatting compliance, phrasing accuracy, and terminology consistency. In a user study (n=14), 90.6% of the generated comments were rated actionable and 67.5% were rated valid, significantly outperforming a GPT-5.2 baseline uswithout the skill library. We release PaperMentor as open source for public use. Our code is publicly available under the AGPL-3.0 license at https://github.com/jiarui-liu/overleaf

cs.CL↗

TamperBench: Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering

As increasingly capable open-weight large language models (LLMs) are deployed, improving their tamper resistance against unsafe modifications, whether accidental or intentional, becomes critical to minimize risks. However, there is no standard approach to evaluate tamper resistance. Varied datasets, metrics, and tampering configurations make it difficult to compare safety, utility, and robustness across different models and defenses. To address this, we introduce TamperBench, the first unified framework to systematically evaluate the tamper resistance of LLMs. TamperBench (i) curates a repository of state-of-the-art weight-space fine-tuning attacks, latent-space representation attacks, and alignment-stage defenses; (ii) enables realistic adversarial evaluation through systematic hyperparameter sweeps per attack-model pair; and (iii) provides both safety and utility evaluations. We use TamperBench to evaluate 21 open-weight LLMs, including defense-augmented variants, across nine tampering threats using standardized safety and capability metrics with hyperparameter sweeps per model-attack pair. The results provide insights including effects of post-training on tamper resistance, that jailbreak-tuning is typically the most severe attack, and that current alignment-stage defenses largely fail to withstand attack sweeps. Code is available at https://github.com/criticalml-uw/TamperBench.

cs.CR↗

Test of Time: Rethinking Temporal Signal of Benchmark Contamination

Post-cutoff performance decay of LLMs has been widely interpreted as a temporal signal for benchmark contamination, where public information released before the training cutoff may have been included into training corpora and inflated model performance by memorization. We critically examine this view and demonstrate that this temporal signal is highly sensitive to how benchmark questions are constructed, even if the underlying source material remains invariant. Specifically, we show that LLM-transformed questions can produce remarkably different temporal patterns compared to fill-in-the-blank (cloze) questions directly retrieved from the very same documents. We validate this effect on prior benchmarks that report clear post-cutoff decay (LiveCodeBench), and show that a simple LLM-driven transformation of the same problems can effectively remove the temporal pattern. We further provide a mechanistic understanding of this phenomenon using influence function analysis. Overall, our results suggest that post-cutoff performance decay is a sensitive contamination signal, motivating more robust contamination probes for reliable LLM evaluation.

cs.AI↗

Accidental Vulnerability: Factors in Fine-Tuning that Shift Model Safeguards

As large language models (LLMs) gain popularity, their vulnerability to adversarial attacks emerges as a primary concern. While fine-tuning models on domain-specific datasets is often employed to improve model performance, it can inadvertently introduce vulnerabilities within the underlying model. In this work, we investigate Accidental Vulnerability, unexpected vulnerabilities arising from characteristics of fine-tuning data. We begin by identifying potential correlation factors such as linguistic features, semantic similarity, and toxicity across multiple experimental datasets. We then evaluate the adversarial robustness of these fine-tuned models, analyzing persona shifts and interpretability traits to understand how dataset factors contribute to attack success rates. Lastly, we explore causal relationships that offer new insights into adversarial defense strategies, highlighting the crucial role of dataset design in preserving model alignment. Our code is available at https://github.com/psyonp/accidental_vulnerability.

cs.CL↗

CORE: Measuring Multi-Agent LLM Interaction Quality under Game-Theoretic Pressures

Game-theoretic interactions between agents with Large Language Models (LLMs) have revealed many emergent capabilities, yet the linguistic diversity of these interactions has not been sufficiently quantified. In this paper, we present the Conversational Robustness Evaluation Score: CORE, a metric to quantify the effectiveness of language use within multi-agent systems across different game-theoretic interactions. CORE integrates measures of cluster entropy, lexical repetition, and semantic similarity, providing a direct lens of dialog quality. We apply CORE to pairwise LLM dialogs across competitive, cooperative, and neutral settings, further grounding our analysis in Zipf's and Heaps' Laws to characterize word frequency distributions and vocabulary growth. Our findings show that cooperative settings exhibit both steeper Zipf distributions and higher Heap exponents, indicating more repetition alongside greater vocabulary expansion. In contrast, competitive interactions display lower Zipf and Heaps exponents, reflecting less repetition and more constrained vocabularies. These results provide new insights into how social incentives influence language adaptation, and highlight CORE as a robust diagnostic for measuring linguistic robustness in multi-agent LLM systems. Our code is available at https://github.com/psyonp/core.

cs.CL↗

SocialHarmBench: Revealing LLM Vulnerabilities to Socially Harmful Requests

Large language models (LLMs) are increasingly deployed in contexts where their failures can have direct sociopolitical consequences. Yet, existing safety benchmarks rarely test vulnerabilities in domains such as political manipulation, propaganda and disinformation generation, or surveillance and information control. We introduce SocialHarmBench, a dataset of 585 prompts spanning 7 sociopolitical categories and 34 countries, designed to surface where LLMs most acutely fail in politically charged contexts. Our evaluations reveal several shortcomings: open-weight models exhibit high vulnerability to harmful compliance, with Mistral-7B reaching attack success rates as high as 97% to 98% in domains such as historical revisionism, propaganda, and political manipulation. Moreover, temporal and geographic analyses show that LLMs are most fragile when confronted with 21st-century or pre-20th-century contexts, and when responding to prompts tied to regions such as Latin America, the USA, and the UK. These findings demonstrate that current safeguards fail to generalize to high-stakes sociopolitical settings, exposing systematic biases and raising concerns about the reliability of LLMs in preserving human rights and democratic values. We share the SocialHarmBench benchmark at https://huggingface.co/datasets/psyonp/SocialHarmBench.

cs.CL↗

Tracing Multilingual Representations in LLMs with Cross-Layer Transcoders

Multilingual Large Language Models (LLMs) can process many languages, yet how they internally represent this diversity remains unclear. Do they form shared multilingual representations with language-specific decoding, and if so, why does performance favor the dominant training language? To address this, we train models on different multilingual mixtures and analyze their internal mechanisms using Cross-Layer Transcoders (CLTs) and Attribution Graphs. Our results reveal multilingual shared representations: the model employs highly similar features across languages, while language-specific decoding emerges in later layers. Training models without English shows identical multilingual shared space structures. Decoding relies partly on a small set of high-frequency features in the final layers, which linearly encode language identity from early layers. Intervening on these features allows one language to be suppressed and another substituted. Finally, to explain non-English failures, we perform a Model-Diffing experiment: underperformance arises from dim late-layer features, weak middle-layer clusters, and tokenizer bias toward English that forces early layers to specialize in word reassembly. Finetuning strengthens these features and their links, improving token assembly and language-specific decoding, providing a mechanistic explanation for multilingual gaps. Our models and CLTs are available at https://huggingface.co/collections/CausalNLP/multilingual-clts and https://huggingface.co/collections/CausalNLP/multilingual-gpt2-models. Our code is available at: https://github.com/abirharrasse/MultilingualCLTs

cs.CL↗

BinaryPPO: Efficient Policy Optimization for Binary Classification

Supervised fine-tuning (SFT) is the standard approach for binary classification tasks such as toxicity detection, factuality verification, and causal inference. However, SFT often performs poorly in real-world settings with label noise, class imbalance, or sparse supervision. We introduce BinaryPPO, an offline reinforcement learning large language model (LLM) framework that reformulates binary classification as a reward maximization problem. Our method leverages a variant of Proximal Policy Optimization (PPO) with a confidence-weighted reward function that penalizes uncertain or incorrect predictions, enabling the model to learn robust decision policies from static datasets without online interaction. Across eight domain-specific benchmarks and multiple models with differing architectures, BinaryPPO improves accuracy by 40-60 percentage points, reaching up to 99%, substantially outperforming supervised baselines. We provide an in-depth analysis of the role of reward shaping, advantage scaling, and policy stability in enabling this improvement. Overall, we demonstrate that confidence-based reward design provides a robust alternative to SFT for binary classification. Our code is available at https://github.com/psyonp/BinaryPPO.

cs.LG↗

Objective Matters: Fine-Tuning Objectives Shape Safety, Robustness, and Persona Drift

Fine-tuning LLMs on benign data can still degrade alignment and adversarial robustness, yet direct analysis of the role of fine-tuning objectives in shaping these safety outcomes remain limited. We present a controlled comparison of six fine-tuning objectives -- Supervised Fine-Tuning, Direct Preference Optimization, Conditional Fine-Tuning, Inoculation Prompting, Odds Ratio Preference Optimization, and KL-regularized fine-tuning -- holding data, domain, architecture, and optimization fixed. Across closed-form reasoning and open-ended generation tasks, we find that objective choice induces systematic, scale-dependent shifts along the safety-capability frontier. At small training budgets, robustness is similar across objectives but capability differs. At larger budgets, objectives diverge sharply: supervised and preference-based tuning tightly couple capability gains to increased adversarial vulnerability and persona drift, while objectives that constrain learning signals -- especially ORPO and KL-regularization -- substantially mitigate both. Fine-tuning objectives therefore matter little for safety at small scales but become a primary driver of adversarial robustness and latent persona stability as training scale increases.

cs.CL↗

Quriosity: Analyzing Human Questioning Behavior and Causal Inquiry through Curiosity-Driven Queries

Recent progress in Large Language Model (LLM) technology has changed our role in interacting with these models. Instead of primarily testing these models with questions we already know answers to, we are now using them for queries where the answers are unknown to us, driven by human curiosity. This shift highlights the growing need to understand curiosity-driven human questions - those that are more complex, open-ended, and reflective of real-world needs. To this end, we present Quriosity, a collection of 13.5K naturally occurring questions from three diverse sources: human-to-search-engine queries, human-to-human interactions, and human-to-LLM conversations. Our comprehensive collection enables a rich understanding of human curiosity across various domains and contexts. Our analysis reveals a significant presence of causal questions (up to 42%) in the dataset, for which we develop an iterative prompt improvement framework to identify all causal queries and examine their unique linguistic properties, cognitive complexity and source distribution. Our paper paves the way for future work on causal question identification and open-ended chatbot interactions. Our code and data are at https://github.com/roberto-ceraolo/quriosity.

cs.CL↗