arXiv subjects

Soroush Vosoughi

Publications and source records attributed to Soroush Vosoughi.

At least 19 recordsLinked to original sources

arXiv2026

Scaling Makes It Possible: How Large Models Master Impossible Languages

While language models (LMs) excel in natural language tasks, their ability to learn impossible languages--those that violate the structural, cognitive, or logical constraints that make human languages learnable and functional--remains debatable. This controversy is further fueled by a divide between linguistic theorists and empirical researchers working with small-scale LMs. To address this controversy, we introduce semantic coherence. This concept measures how effectively semantic information is expressed in language structures and examines the learning capabilities of LMs across multiple scales. By fine-tuning billion-parameter LMs on impossible language samples, we demonstrate that larger LMs achieve similar performance on impossible and possible languages, whereas million-parameter LMs exhibit significant degradation when trained on impossible languages. This performance gap widens as semantic complexity increases. We further show that scaling up the LM size fundamentally shifts the learning focus from grammatical rules to semantic understanding, with billion-scale LMs exhibiting a threefold reduction in sensitivity to low-level linguistic perturbations while maintaining stable performance under high-level semantic perturbations.

cs.CL↗

arXiv2026

Priority-Preserving Advantage Aggregation for Multi-Reward Policy Optimization

LLM post-training often optimizes multiple objectives with unequal priorities, such as correctness and reasoning efficiency. However, the widely used additive reward aggregation in RL can violate these priorities: when primary and auxiliary advantages conflict, a sufficiently large auxiliary signal can overturn the preference induced by the primary objective. We term this problem \emph{advantage priority inversion}. Although reward-level gating can protect priority more directly, it still removes potentially useful auxiliary distinctions among rollout samples that fail the primary condition. To address this trade-off, we propose Primary Advantage Policy Optimization (\textbf{PAPO}), a priority-aware advantage aggregation method that asymmetrically combines a primary and an auxiliary advantage. PAPO retains auxiliary supervision when the objectives agree and smoothly attenuates conflicting auxiliary signals, preserving the primary sign during fusion for \emph{primary-consistent ordering}. This rule applies to sequence-level advantages in GRPO and token-level advantages in actor--critic PPO. Experiments on multiple domains and models show that PAPO consistently achieves the best primary objective while maintaining best or competitive auxiliary-objective performance across both GRPO- and PPO-based optimization. \href{https://xiangchi-yuan.github.io/papo-project/}{Code is available}.

cs.LG↗

arXiv2026

Beyond Belief: Evaluating and Mitigating LLM Susceptibility to Misinformation

Tool-using large language models (LLMs) must decide whether external information warrants action. We investigate whether their expressed credibility judgments reliably indicate behavioral resistance to misinformation. We evaluate five LLMs on 196 constructed false notices paired with legitimate counterparts and 204 archived attacks adapted to simulated tool use. When asked to act on the constructed notices, models issue harmful tool calls in 27%--99% of cases they have rated untrustworthy. Removing the preliminary rating step increases harmful execution by 15 percentage points on average, showing that credibility elicitation itself changes the behavior being evaluated. Interventions that improve credibility discrimination do not consistently yield safer actions. A simple forewarning significantly reduces harmful execution in four of five models, but neither tested generated-example prebunk outperforms it in the main evaluation. These reductions can also compromise legitimate task completion. Supervised fine-tuning on generated false notices reduces harmful execution on held-out archived attacks under the rating-first protocol, with model-dependent utility costs and weaker protection without rating elicitation. Our findings motivate evaluating misinformation defenses through both harmful actions and legitimate task completion, while accounting for the behavioral effects of the evaluation protocol.

cs.CL↗

arXiv2026

Target-Aware Curation of Hallucination Stress Tests for Large Language Models

Hallucination stress tests can yield similar average scores across language models even when different prompts elicit hallucinations in each model. We develop a procedure that iteratively edits source-grounded seed prompts to expose hallucinations in a target model. An LLM editor revises each prompt in response to the target's answer and a judge's category-specific score and explanation. In the main setting, the target also receives feedback on its previous response. Experiments cover 500 seeds from seven datasets and ten topics, with intrinsic and extrinsic hallucinations evaluated across open-weight and closed-source models. Editing substantially increases judge-rated severity over the seeds. In the baseline configuration, retained responses fall in the judge's strong-hallucination bin in 77.2% of runs targeting extrinsic hallucination and 74.2% targeting intrinsic hallucination. Later editing rounds account for more than half of the mean score increase. Among 219 prompts that pass model-assisted source-validity checks, 89 elicit responses containing unsupported claims. Across five models, 200 curated prompts yield similar average severity scores, although pairwise Spearman correlations between prompt-level scores are only $ρ=0.0387$ to $0.1897$. Curated prompts can be reused across models, but their diagnostic value depends on source validity and the failures they expose in each model.

cs.CL↗

arXiv2026

Of Errors and Echoes: Responsibility-Aware Credit Assignment for Tool-Use Agents

Reinforcement learning improves tool-use agents, yet fine-grained credit assignment remains challenging even with step-level feedback. An observed error may be inherited from an earlier decision, while multiple responsible decisions may share the same downstream consequence or participate in later recovery. We introduce Error-Chain-Aware Optimization (ECHO), a responsibility-aware credit assignment framework that separates responsibility identification from credit allocation. A frozen judge combines retrospective trajectory analysis with decision-time assessment from the policy-visible prefix to distinguish newly introduced errors, inherited consequences, missed corrections, and successful corrections. ECHO then counts shared downstream consequences once across responsible decisions and constrains rewards for repairing self-induced errors through source-funded recovery, producing bounded action-level supervision alongside terminal supervision. Experiments on six diverse tool-use benchmarks at both 4B and 8B scales show strong and consistent improvements over the compared baselines, with ECHO achieving the highest score on five of six benchmarks at each scale. These results support decision responsibility as a useful basis for fine-grained credit assignment in tool-use agents. Our code is available at https://github.com/xid32/ECHO.

cs.CL↗

arXiv2026

PRISM-BN: A Controlled Corpus and Benchmark for Text-to-Parameterized Bayesian Network Extraction

Probabilistic Graphical Models (PGMs), especially Bayesian Networks (BNs), expose directed structure and probabilistic parameters, making them natural symbolic targets for neurosymbolic AI. Yet training text-to-parameterized-BN systems requires paired text-to-BN resources unavailable at scale. We introduce PRISM-BN, a controlled corpus of 5054 BN-grounded descriptions paired with discrete reference BNs containing variables, states, directed edges, root priors, and full multi-parent CPDs across five domains. The instances are derived from 50 Wikipedia-seeded backbones, and their probabilities are internally constructed benchmark targets rather than externally validated causal estimates. PRISM-BN is built with PRISM, a marginal-first pipeline that elicits marginal and local joint distributions, analytically recovers normalized CPDs, and constructs locally reparameterized subgraphs. We define a benchmark with semantic node and state alignment, conditional structural scoring, and strict full-CPD evaluation. Across six LLM extractors, Node F1 ranges from 0.56 to 0.83, conditional Edge F1 from 0.90 to 0.97, and CPD-KL from 1.11 to 3.14. Conditional state and edge recovery remain consistently strong, whereas strict full-CPD agreement remains challenging. These trends persist with independently generated GPT-5.5 references, and a human pilot corroborates structural recoverability and similar probabilistic interpretations. PRISM-BN supports separate evaluation of structural recovery and probabilistic parameter estimation.

cs.CL↗

arXiv2026

Which Tokens Should SFT Actually Learn? A Token-Trimming Perspective on Mathematical Reasoning

Supervised fine-tuning (SFT) applies a uniform cross-entropy loss to all target tokens, even though different tokens provide unequal learning signals for mathematical reasoning. This uniform treatment can over-sharpen already mastered tokens while amplifying learning pressure on uncertain, low-confidence tokens, leading to suboptimal training dynamics. We propose Trimmed Logit-Gap SFT (TrimSFT), a simple token-level reweighting method that scales the SFT loss according to the logit gap between the gold token and its strongest competitor. TrimSFT trims supervision away from both extremes: tokens already mastered (large logit gap) and tokens weakly supported by the current model (small or negative logit gap), concentrating learning within an intermediate logit-gap region between them. We instantiate this principle with a Gaussian weight centered at margin m with bandwidth τ, requiring no reference model or additional forward pass. We evaluate TrimSFT on six base models from the Llama, Qwen, and DeepMath families across five mathematical reasoning benchmarks. TrimSFT consistently improves over standard SFT, achieving the best average performance on five out of six models, with gains of up to +26.9 points over SFT on MATH500. Further analyses show that the bandwidth τ matters more than the exact margin location, and that half-trim variants that remove supervision pressure from only one side yield inferior trade-offs. A token-level logit-gap distribution analysis suggests that TrimSFT reshapes model confidence in a more balanced way than uniform SFT or monotonic reweighting methods. These results suggest that reasoning SFT can benefit from trimming both extremes rather than treating all tokens uniformly.

cs.AI↗

arXiv2026

Kernel-Complexity Edge Sanitization for Training-Free Defense against Structural Graph Attacks

Graph Neural Networks (GNNs) have achieved remarkable success across diverse applications, yet they remain highly vulnerable to adversarial attacks that maliciously perturb graph structure. Existing defenses often lack rigorous theoretical grounding, rely on attack-specific heuristics, or require costly retraining procedures such as adversarial training. To address these limitations, we propose Kernel-Complexity Edge Sanitization (KCES), a training-free and model-agnostic framework for defending against structural attacks. KCES is built upon Graph Kernel Complexity (GKC), a principled metric derived from the graph Gram matrix that appears in a generalization upper bound on the GNN test error. From this bound, we define an edge-specific KC score that quantifies each edge's structural influence via its induced change in GKC. KCES then identifies and prunes high-KC edges, which are empirically enriched with adversarial perturbations under structural attacks, to mitigate their harmful impact. Computationally efficient and scalable, KCES operates as a lightweight preprocessing step without retraining and can be seamlessly integrated with existing defenses. Extensive experiments demonstrate that KCES consistently outperforms representative robust baselines across diverse attack settings and scales effectively to large graphs. Supported by theoretical analysis and extensive empirical validation, KCES provides a principled and efficient framework for securing GNNs. Our code is available at https://github.com/karpning/KCScore.

cs.LG↗

arXiv2026

When Linguistic and Internal Confidence Diverge in Large Language Models

Users often ask large language models (LLMs) to report how confident they are, but it is unclear whether such linguistic confidence tracks the model's internal confidence. We study this question across 8 classification tasks, 2 generation tasks and 30 models from three families. For classification, we compare linguistic confidence with logits-based confidence along three axes: association, magnitude agreement and calibration. For generation, we test whether linguistic confidence tracks semantic-entropy-based uncertainty. The axes frequently diverge. Instance-level association is weak on average, although it improves on easier items and for stronger base models. Instruction-tuned models often report higher confidence and sometimes show higher association, but they also have larger confidence gaps and worse calibration. Prompt design mostly changes the distribution of reported confidence. Attitude cues inflate confidence without improving alignment, while score exemplars can preserve rank-order signal when they avoid collapsed confidence values. Regression analyses show that distributional properties of confidence scores explain much of the observed alignment pattern, with model metadata playing a smaller role after controls. These results support a lossy-channel view of linguistic confidence. A more dispersed verbal confidence distribution can carry useful rank information, but it does not make the scores calibrated. Linguistic confidence should therefore be evaluated with multi-axis diagnostics before being used in downstream reliability pipelines.

cs.CL↗

arXiv2026

UTILMEM: Benchmarking Evidence Utilization in Long-Term Conversational Memory

Long-term memory is increasingly important for conversational agents, yet existing benchmarks primarily measure memory through pointwise factual recall: whether a system can recover isolated facts or event-level details from prior interactions. Real-world memory use, however, often requires a more demanding capability: integrating distributed, implicit, and noisy evidence across extended interaction histories into coherent, task-oriented outputs. We call this capability memory utilization. Here, we introduce UtilMem, a diagnostic benchmark comprising 1,717 instances across five domains, designed to evaluate four underexplored aspects of memory utilization: reasoning over dense histories, identifying implicitly relevant memories, synthesizing distributed evidence into summaries, analyses, or plans, and resisting interference from semantically similar distractors. Evaluating a diverse set of retrieval-based and memory-augmented systems, we find that strong performance on conventional factual-memory benchmarks does not reliably translate into effective memory utilization. Moreover, retrieval alone is insufficient: even when relevant evidence is successfully recovered, systems frequently fail to integrate information across sessions or to distinguish useful evidence from plausible distractors. These findings expose a substantial gap between accessing stored information and using it effectively, and suggest that progress in long-term conversational memory will require architectures that explicitly support evidence integration and robustness to retrieval interference. Code is available at https://github.com/peijunallin/UtilMem.

cs.CL↗

arXiv2026

Improving Information Extraction with Learned Queries

When information extraction fails, a natural instinct is to improve the model doing it: for example, by scaling it up or refining its reasoning. In this paper, we show that another part of the pipeline matters at least as much: the queries used to elicit this information. Across four clinical benchmarks and five LLMs, improving the question design alone raises performance by 18.6 F1-score points, i.e. more than using larger extraction models. To make such question design learnable, we introduce List of Questions (LoQ), which generates document-specific question sets, and FeedQ, a feedback-driven optimization method that iteratively refines questions against extraction outcomes. The resulting optimized questions can be used to train lightweight generators: with fine-tuning, 4B-parameter models match or outperform expert-derived baselines and substantially exceed the performance of much larger untuned models. We release a dataset of 12,820 optimized questions to support a broader shift in information extraction research toward treating question design as a first-class problem.

cs.CL↗

arXiv2026

Scalable and Culturally Specific Stereotype Dataset Construction via Human-LLM Collaboration

Research on stereotypes in large language models (LLMs) has largely focused on English-speaking contexts, due to the lack of datasets in other languages and the high cost of manual annotation in underrepresented cultures. To address this gap, we introduce a cost-efficient human-LLM collaborative annotation framework and apply it to construct EspanStereo, a Spanish-language stereotype dataset spanning multiple Spanish-speaking countries across Europe and Latin America. EspanStereo captures both well-documented stereotypes from prior literature and culturally specific biases absent from English-centric resources. Using LLMs to generate candidate stereotypes and in-culture annotators to validate them, we demonstrate the framework's effectiveness in identifying nuanced, region-specific biases. Our evaluation of Spanish-supporting LLMs using EspanStereo reveals significant variation in stereotypical behavior across countries, highlighting the need for more culturally grounded assessments. Beyond Spanish, our framework is adaptable to other languages and regions, offering a scalable path toward multilingual stereotype benchmarks. This work broadens the scope of stereotype analysis in LLMs and lays the groundwork for comprehensive cross-cultural bias evaluation.

cs.CL↗

arXiv2026

When Debiasing Backfires: Counterintuitive Side Effects of Preprocessing-Based Stereotype Mitigation

Preprocessing-based methods for stereotype mitigation, such as pre-/post-training on debiased corpora, are widely used in NLP. While these approaches reduce measurable stereotypes for targeted groups, we find they often induce unintended shifts-side effects, where stereotyping or counter-stereotyping can increase relative to neutral baselines for other demographics, including across unrelated demographic categories. We demonstrate these side effects across two model families (encoder-only and decoder-only), multiple preprocessing strategies (removing stereotypical sentences, removing group mentions, and swapping group references), and both pre- and post-training at different data scales on Wikipedia. Standard benchmarks frequently miss these shifts. Using attention-rollout analysis, we observe that such side effects are not accompanied by large changes in attention flow, complicating mechanistic explanations. We discuss implications for evaluation, provide actionable diagnostics, and argue for side-effect-aware, transparent mitigation practices.

cs.CL↗

arXiv2026

Memory Makes the Difference: Evaluating How Different Memory Roles Shape Conversational Agents

Prior research on memory mechanism in RAG-based conversational system has emphasized how memory is stored and retrieved. However, far less is known about how memories with different functional roles influence response quality. Specifically, how they shape an agent's responses under varying conversational contexts and whether they lead to substantively different response behaviors. Existing evaluations in conversational system are also largely reference-based, insufficiently capturing the nuances in responses that may address users' preferences differently. In this work, we probe the impact of different memory types in shaping agents' responses. We present a fine-grained taxonomy of conversational memory, classify retrieved memories into different role types, and design a user-centric evaluation framework that simulates user perspectives. Through comparative experiments on long-term datasets and frontier LLMs, our analysis reveal many differentiated effects of memories: e.g., clarifying memory improves responses' factual accuracy and constraint awareness, making them more correct and personalized; irrelevant memory reduces topic relevance and degrades constraint awareness. Despite the power of frontier LLMs, these findings shed light on how different memory types can be leveraged to produce more personalized responses and inspire further research in this direction.

cs.CL↗

arXiv2026

The Hidden Signal of Verifier Strictness: Controlling and Improving Step-Wise Verification via Selective Latent Steering

Generative verifiers have emerged as a promising paradigm for step-wise verification, but their verification behavior is often poorly calibrated: they may be under-critical and miss erroneous steps, or over-critical and reject correct reasoning. We refer to this tendency to be overly lenient or overly critical as verifier strictness. In this work, we study whether verifier strictness can be controlled through hidden-state intervention. We uncover a verification-specific hidden-state signal: in step-wise verification, a verifier's tendency to accept or reject a solution step is encoded near the boundary of the corresponding verification paragraph. Exploiting this signal, we show that hidden-state steering can directly modulate verifier strictness without fine-tuning. However, uniform steering induces a trade-off between error detection and correctness certification. To address this, we propose VerifySteer, which exploits latent correctness signals for sample-level routing and selectively intervenes on paragraph boundaries. Experiments on ProcessBench and Hard2Verify show that VerifySteer outperforms prompt optimization and activation steering baselines, and is competitive with self-consistency while requiring 4-7x less inference compute. VerifySteer is also complementary to verification fine-tuning, providing further gains on top of fine-tuned verifiers. The code is available at https://github.com/YefanZhou/VerifySteer.

cs.LG↗

arXiv2026

Overcoming Multi-step Complexity in Multimodal Theory-of-Mind Reasoning: A Scalable Bayesian Planner

Theory-of-Mind (ToM) enables humans to infer mental states-such as beliefs, desires, and intentions-forming the foundation of social cognition. However, existing computational ToM methods rely on structured workflows with ToM-specific priors or deep model fine-tuning, which struggle with scalability in multimodal environments and fail to generalize as task complexity increases. To address these limitations, we propose a scalable Bayesian ToM planner that decomposes ToM reasoning into stepwise Bayesian updates. Our framework introduces weak-to-strong control, allowing smaller language models (LMs) to specialize in ToM-specific likelihood estimation and transfer their reasoning behaviors to larger LMs (7B to 405B) for integration with social and world knowledge. This synergistic approach aligns large-model inference of human mental states with Bayesian principles. Extensive experiments show that our method achieves a 4.6% accuracy improvement over state-of-the-art techniques on multimodal ToM benchmarks, including challenging unseen scenarios, thereby establishing a new standard for modeling human mental states in complex environments.

cs.AI↗

arXiv2026

What Makes LLMs Effective Sequential Recommenders? A Study on Preference Intensity and Temporal Context

What enables large language models (LLMs) to effectively model user preferences in sequential recommendation? Our investigation reveals that existing preference-alignment approaches largely rely on binary pairwise comparisons, overlooking two critical factors: preference intensity (the structured strength of affinity or aversion) and temporal context (the extent to which recent interactions better reflect a user's current intent). Through controlled experiments, we show that leveraging comprehensive feedback with structured preference signals substantially improves recommendation performance, indicating that binary modeling discards essential information. Motivated by these findings, we propose RecPO, a unified preference optimization framework that maps both explicit and implicit feedback into a common preference signal and constructs adaptive reward margins that jointly account for preference intensity and interaction recency. Experiments across five datasets show that RecPO consistently outperforms state-of-the-art baselines while exhibiting behavioral patterns aligned with human decision-making, including favoring immediate satisfaction, maintaining preference coherence, and avoiding dispreferred items. Our results highlight that preference intensity and temporal context are fundamental ingredients for effective LLM-based recommendation.

cs.IR↗

arXiv2026

Addressing Overthinking in Large Vision-Language Models via Gated Perception-Reasoning Optimization

Large Vision-Language Models (LVLMs) have exhibited strong reasoning capabilities through chain-of-thought mechanisms that generate step-by-step rationales. However, such slow-thinking approaches often lead to overthinking, where models produce excessively verbose responses even for simple queries, resulting in test-time inefficiency and even degraded accuracy. Prior work has attempted to mitigate this issue via adaptive reasoning strategies, but these methods largely overlook a fundamental bottleneck: visual perception failures. We argue that stable reasoning critically depends on low-level visual grounding, and that reasoning errors often originate from imperfect perception rather than insufficient deliberation. To address this limitation, we propose Gated Perception-Reasoning Optimization (GPRO), a meta-reasoning controller that dynamically routes computation among three decision paths at each generation step: a lightweight fast path, a slow perception path for re-examining visual inputs, and a slow reasoning path for internal self-reflection. To learn this distinction, we derive large-scale failure attribution supervision from approximately 790k samples, using teacher models to distinguish perceptual hallucinations from reasoning errors. We then train the controller with multi-objective reinforcement learning to optimize the trade-off between task accuracy and computational cost under uncertainty. Experiments on five benchmarks demonstrate that GPRO substantially improves both accuracy and efficiency, outperforming recent slow-thinking methods while generating significantly shorter responses.

cs.CV↗