Skip to content
AI NEWS JOURNAL
AI, IN PERSPECTIVE

Search AI news

Search headlines, summaries and articles to connect the dots.

Newest first

EgoLink: Egocentric Language-Vision Interactive Network Knowledge Challenge Summary

arXiv:2611.21690v1 Announce Type: new Abstract: The EgoLink Challenge at ACM Multimedia 2026 benchmarks egocentric multimodal intelligence linking first-person perception, social reasoning, and interactive action. Unlike benchmarks centered on visible objects and activities, EgoLink evaluates whether models can interpret emotional and social cues, resolve visually grounded references, and convert user requests into verified tool-mediated actions. Its two tracks cover social reasoning in egocentric video and interactive agents that complete service tasks through dialogue, tool use, and closed-loop execution. Together, they test whether perception, reasoning, and decision-making remain coupled under both offline question answering and state-changing interaction. We highlight a multi-stage Track 1 data pipeline for multiple-choice question (MCQ) generation, chain-of-thought generation, automatic quality grading, and video-contribution filtering. We summarize the task design, evaluation metrics, rules, leaderboards, and representative solutions. The benchmark combines objective answer scoring with process- and result-based evaluation, enabling analysis of both final correctness and execution quality. The best Track 1 system achieved 0.9263 accuracy, while the best Track 2 system achieved 0.844 joint success and 0.916 micro accuracy. Participant reports reveal a shared principle: strong systems make intermediate evidence explicit through candidate verification, option-robust reasoning, scenario-aware planning, tool canonicalization, and state-aware correction before answering or acting. These findings highlight the importance of process verification and evidence-grounded reasoning for reliable egocentric multimodal intelligence.

FieldGuide: A Corpus Map for Token-Efficient Agentic Document Exploration

arXiv:2611.21647v1 Announce Type: new Abstract: Generative AI agents can help journalists explore large document collections. But exploratory workflows can require substantial token use and time, especially when the task is broad and the relevant subquestions are not yet known. This paper introduces FieldGuide, an agent-facing corpus map that organizes a document collection into a topic hierarchy and exposes tools for corpus orientation, scoped search, and bounded reading. I evaluate FieldGuide on five journalistic document-exploration tasks using 38 Chicago Police Department public-records documents, comparing it with raw-file access and retrieval-augmented generation. Across the five tasks, FieldGuide used 27% fewer tokens and 41% less elapsed time than raw-file access; relative to RAG, it used 38% fewer tokens and 21% less elapsed time. An accuracy audit identified four minor issues in FieldGuide's responses, comparable to the other conditions. In blind ratings of generated reporting leads, FieldGuide's mean ratings were higher than RAG's and lower than the raw-file condition's. These descriptive results suggest that corpus orientation may reduce runtime token use and elapsed time without severe accuracy failures. They also motivate further study of whether selective access limits broad, cross-document discovery.

Flux: Optimal Scheduling of Optical Circuit Switches for LLM Training

arXiv:2609.25949v2 Announce Type: replace-cross Abstract: Optical Circuit Switching (OCS) offers high bandwidth density and energy efficiency for LLM training, but incurs a non-negligible reconfiguration delay. Prior work typically schedules optical circuit switches independently of compute, using aggregate traffic demand to determine which circuits to provision and when. We argue that this separation creates a fundamental inefficiency: reconfigurations that ignore the compute timeline can stall communication, resulting in low circuit utilization and large buffer requirements. In this paper, we present Flux, a scheduler that optimally schedules optical circuit switches based on the structure of the entire workload. Flux remains effective across a wide range of switching speeds by reusing circuits and amortizing reconfiguration delay behind compute and communication. We show that Flux reduces training iteration time by up to $10\times$ and peak NIC buffer requirements by more than three orders of magnitude compared to traditional periodic schedulers.

From Preference Data to Personalization: Tracing Sycophancy in Large Language Models

arXiv:2611.21650v1 Announce Type: new Abstract: Sycophancy in large language models is often measured as agreement with the user, yet agreement can be appropriate when the user's claim is valid, their preference is genuine, or their experience warrants acknowledgment. We study sycophancy as inappropriate agreement: validating a claim, preference, or self-assessment when the context calls for correction, qualification, or resistance. We introduce a validity-aware sycophancy grader and SycoDrift, a persona-targeted evaluation set of 3,298 validity-annotated probes with controlled annotations for user openness and stakes. We use this framework to analyze three sources of agreement pressure: human preference data, automated judge specification, and in-context personalization. First, preference for validating responses in human datasets (Chatbot Arena, HH-RLHF, Community Alignment) is weak and concentrated where agreement is often appropriate. Second, in a selected-value, single-model experiment, emphasized judge rubrics produce consistent but modest increases relative to paired suppressed rubrics, demonstrating a plausible mechanism for training-time amplification. Third, in off-policy evaluations across 14 models, inference-time personalization increases inappropriate agreement, with the largest effect from matched user profiles placed directly before the query.

JigSLO: Solving the Resource Puzzle for Efficient Multi-LLM Serving

arXiv:2509.06261v3 Announce Type: replace Abstract: GPU spatial sharing co-locates instances with complementary demands, such as compute-bound prefill and memory-bound decode, promising cost-efficient multi-LLM serving. Prior systems partition compute and coordinate shared memory capacity, but do not model bandwidth contention among co-located instances of heterogeneous LLMs. This contention makes each instance's latency depend on the memory-access patterns of its co-runners, even when its own resource allocation remains unchanged. Drawing on extensive measurements, we find that the latency impact of bandwidth contention can be captured by a time-shared memory-access model with contended and uncontended periods, so the overall latency can be analytically estimated as a time-weighted average of the latency estimates for these periods. We propose JigSLO, a contention-aware multi-LLM serving system that manages memory resources and schedules each model's prefill and decode phases as separate instances, treating their co-location as a placement decision. Its contention-aware predictor estimates per-instance latency under any co-location from standalone profiles alone. Guided by these predictions, its placement algorithm jointly selects each instance's tensor parallelism, compute share, and batch configuration, pairing complementary memory-access patterns to maximize throughput within each model's latency targets. JigSLO's shared resource manager complements this on the capacity side, sharing weights, pooling KV cache, and transferring it between prefill and decode instances. Across three datasets covering chat, code generation, and long-context summarization, JigSLO outperforms state-of-the-art baselines by $2.3\times$ in SLO attainment and $1.2\times$ in throughput on average, matching their best attainment with half the GPUs on long-context workloads.

KV-LTM: Persistent Per-Session KV Caches as Long-Term Memory for Personal Assistants

arXiv:2611.21998v1 Announce Type: new Abstract: Personal assistants accumulate months of conversations, but today's long-term memory stacks put derivation or selection between a question and its evidence. In a system-level comparison on LongMemEval, two commercial assistants scored well below the same model reading the history offline [1], consistent with, but not isolating, content loss between write and read. We propose KV-LTM: a memory tier storing each session's raw text plus its KV cache under a fixed reader model, answering by restoring cached sessions and appending the question. KV-LTM is complementary to, not a replacement for, retrieval and indexing. Text is the memory of record; the cache is a recompute-free prefill accelerator, degrading gracefully to text. We specify the design: storage record, budget-derived retention, append-only read path, and bounded aggregation. We size it analytically: across pinned BF16 full-attention (GQA) and latent-cached MLA configurations, per-token state spans approximately 31K-328K bytes (the pinned KDA+MLA hybrid differs: approximately 8.1K bytes/token plus a fixed per-session recurrent term), giving approximately 0.34-3.6 TB per user-year at a declared 30K-token/day scenario, against approximately 49 MB of raw text; per-query cost is modelled as regime-conditional transfer, attention, decode, aggregation, and compute. We define an evaluation protocol needing no persistent-cache implementation for logical QA evaluation: per-session prefill emulates valid cache reuse under the same model/input-state contract and ideal arithmetic, with named arms, controlled baselines, and a preregistered reader-model gate; a separate minimal real-cache replay tests numerical equivalence and measures the cost axis, with outcomes not assumed. KV-LTM trades storage for avoided re-prefill compute and write-time fidelity, within a declared retention budget.

Latent Fact Composition under Next-Token Training

arXiv:2611.21921v1 Announce Type: new Abstract: Controlled studies of latent two-hop reasoning train transformers to predict only the answer, whereas language models are trained to predict every token. We ask whether a model trained on every token learns to answer two-hop questions whose intermediate entity appears in no two-hop training example. We build a synthetic corpus with such questions. On this corpus every term of the loss can be computed exactly. A looped transformer, a standard transformer and DiscoLoop answer almost none of these questions. We therefore introduce the Baton family of looped transformers. They pass the intermediate entity to the second loop as a token embedding. When trained to predict only the answer, two of its three variants answer these questions almost perfectly. We then relate the success and failure of each model to three design conditions. Experiments that change one condition at a time agree with this hypothesis. When Baton is trained on every token, however, its accuracy on these questions rises and then declines. We further split the gradient of the next-token loss into three terms. Removing the two terms that depend on the particular training set raises the peak accuracy but does not remove the decline. The decline arises in the second hop while the model still recalls both facts.

MInfoVisQA: A Multilingual Benchmark for Information-Rich Visual Question Answering

arXiv:2611.21644v1 Announce Type: new Abstract: Multimodal large language models (MLLMs) excel in vision-language tasks; yet, whether they maintain consistent understanding of information-rich visuals (InfoVis) when the surface language changes remains unclear. To quantify this capability, we introduce MInfoVisQA, a parallel benchmark with 8,960 QA configurations derived from 128 seed instances across 24 languages. We construct seed-matched, language-parallel visuals and QA pairs by disentangling numerical-structural data from text and programmatically synchronizing translated labels. We evaluate 25 MLLMs under two settings: Language-Aligned QA (LQA, matching visual-text languages) and Cross-Language QA (XQA, fixed QA language with varying visual languages). The results show that aggregate accuracy masks substantial robustness gaps: two models with comparable mean LQA accuracy (80.4\%) exhibit performance spreads of 7.8 and 20.3 percentage points across languages. Moreover, all 25 models score lower with non-English visuals under fixed English QA, underscoring the need for cross-language sensitivity metrics.

NEBench R1: a Reasoning Benchmark for Evaluating Nuclear Domain Competence in Large Language Models

arXiv:2611.21682v1 Announce Type: new Abstract: Large language models (LLMs) are increasingly being explored for knowledge-intensive and safety-critical applications in nuclear engineering. However, existing nuclear-domain benchmarks provide limited content coverage or focus primarily on answer accuracy without systematically evaluating the quality of the underlying reasoning. This study introduces NEBench R1, a reasoning-oriented benchmark comprising 7,546 unique multiple-choice questions extracted from U.S. Nuclear Regulatory Commission (NRC) reactor operator licensing examinations covering 18 nuclear facilities. Each question is classified according to the NRC Knowledge and Abilities catalogs and accompanied by an official explanation serving as a reference for reasoning evaluation. Six LLMs were evaluated, with overall accuracy ranging from 34.0% to 73.9% across three distinct performance tiers. To assess reasoning quality, an LLM-based judge evaluated correctly answered responses using a multidimensional rubric. Although responses generally demonstrated strong problem comprehension, they scored lower in factual and procedural accuracy and distractor analysis, indicating that correct answers were not necessarily supported by technically sound justifications. Response clarity was also weakly indicative of technical quality. Furthermore, longer reasoning traces were associated with lower accuracy in both frontier reasoning models, suggesting that reasoning length may reflect question difficulty rather than improved problem-solving performance. Reasoning-token consumption may therefore provide an additional signal for identifying potentially unreliable responses. NEBench R1 establishes a benchmark for evaluating both answer accuracy and reasoning quality, providing a foundation for assessing the domain competence and reliability of LLMs in nuclear engineering.

On Formal Counterfactual Explanations

arXiv:2611.22010v1 Announce Type: new Abstract: Counterfactual explanations are often seen as counterfactual conditionals under closest-world semantics, requiring one to reason over all minimally distant counterexamples. In contrast, formal XAI typically relies on feature-based abductive explanations (AXps) and contrastive explanations (CXps), which admit efficient computation and a minimal hitting-set duality. This paper bridges these perspectives by (i) characterizing closest-world counterfactual explanations explicitly as a disjunctive normal form over all minimally distant counterexample instances, and (ii) introducing conjunctive counterfactual explanations (CFXps), a subclass of counterfactual antecedents restricted to conjunctions of feature literals that guarantee prediction change. The paper establishes structural results connecting CXps, AXps and CFXps: every CXp induces an AXp for its witness instance, and every CXp together with a witness AXp induces a CFXp whose support is precisely determined by that AXp. These theoretical results yield a modular approach for computing CFXps using existing AXp/CXp machinery. Experiments on a representative set of benchmarks using tree ensembles and neural networks show that CFXp extraction is practically efficient. Compared with popular point-based counterfactual explainers, CFXps generally require no more feature changes and can guarantee the counterfactual prediction beyond a single instance.

PromptPack: Training-Free Augmentation of Diffusion Editors for Intent-Aligned Spatial and Semantic Control

arXiv:2611.21865v1 Announce Type: new Abstract: Many image-editing tasks require confining changes to intended spatial regions while exactly preserving user-designated source content. Instruction-driven editing pipelines often leave such user-intent underspecified, while inpainting editors condition generation on explicitly defined edit support but not on the spatial assignment of individual edit concepts. PromptPack addresses both limitations by equipping diffusion editors with two explicit 'Edit Guides': an edit-role partition that labels each pixel as 'Keep,' 'Permit,' or 'Modify' to specify where edits may occur, and a regional-prompt assignment that maps spatial regions to text prompts to specify which concepts apply where. These guides may be manually specified or, for source-object edits, automatically derived from a text-grounded SAM 3 mask. Without retraining, PromptPack injects independently embedded regional prompts into the editor's fixed-length text sequence and uses the Edit Guides to spatially mask attention, gate flow updates, and restore protected source pixels. We integrate PromptPack into FlowEdit FLUX, FlowEdit SD3, and FLUX Kontext and demonstrate improved requested-concept placement and protected-content preservation across all three integrations on PIE-Bench++. In a pairwise user study, PromptPack achieves an 87% preference rate over FlowEdit FLUX, establishing PromptPack as a practical, training-free interface for intent-aligned diffusion editing.

RAW: Refusal Aversion in the Wild, A Causal Measurement Method for Deployed LLMs

arXiv:2611.21641v1 Announce Type: new Abstract: Deployed LLMs field hundreds of millions of conversations a day and refuse a non-trivial fraction of user requests, yet the user-side cost of refusal, whether users disengage and by how much, is unmeasured on any production model: a blind spot for the consumer-facing platforms built on them. A/B-testing refusal is ethically and operationally constrained on frontier-lab products and unavailable to third-party platforms deploying these models through a fixed API. We introduce RAW, a method that uses LLM stochasticity as a natural experiment: at near-identical prompts in any existing log, the same model sometimes refuses and sometimes does not, and that variation identifies the causal user-side effect of refusal. Applying the method to 15 deployed LLMs in LMSYS-Chat-1M and WildChat, with user-clustered inference, we find that for about half of the models we can measure precisely, a refusal costs 3 to 17 pp of subsequent re-engagement; the rest show no distinguishable effect, among them one deployment whose positive point estimate we trace to persistent users retrying the refused request. Our contribution is the measurement method, which runs on any deployed model from logs alone for about \$0.30 per 1,000 conversations. We release labels, moderation scores, prompt embeddings, and pipeline code.

REEL: Brain-Guided Video Generation and Selection by Reranking Encoder Embeddings with Linear Decoding from fMRI data

arXiv:2611.21907v1 Announce Type: new Abstract: Visual brain decoding from functional magnetic resonance imaging (fMRI) data has made remarkable progress in recent years, driven by increasingly powerful artificial intelligence (AI) models and the growing availability of large-scale public datasets. However, despite the impressive advances achieved in static image decoding, the reconstruction of videos from fMRI signals remains challenging due to the temporal mismatch between rapidly changing video stimuli and the slow dynamics of the blood-oxygen-level-dependent (BOLD) response. We evaluate a Bayesian-inspired proposal-and-rerank strategy on CC2017 and the BOLD Moments Dataset (BMD), using a linear semantic decoder, a pretrained video generator, and three neural encoding streams with ridge readouts and learned pooling in two streams. We refer to this framework as REEL (Reranking Encoder Embeddings with Linear decoding). Compared with a single generated candidate, reranking increases mean Video 50-way accuracy from 0.2359 to 0.2434 across three CC2017 participants and from 0.3653 to 0.3920 across ten BMD participants. These findings suggest an alternative and conceptually simpler strategy for improving video brain decoding, highlighting the potential of our approach as a complementary direction to increasingly complex pipelines.

Regularized Score Centering

arXiv:2611.21791v1 Announce Type: new Abstract: Asynchronous reinforcement learning increases training throughput, but data from stale policies and numerical mismatch between training and inference can destabilize learning. To stabilize off-policy learning, score centering (SC) removes the effect of reward baselines on the expected update. Our gradient diagnostics show less extreme gradients with SC than with importance sampling (IS), but SC can still fail when the learner moves far from the behavior policy that generated the data. To limit this policy mismatch, we introduce Regularized Score Centering (RSC), which adds a reverse Kullback-Leibler (KL) penalty to encourage the learner's next-token distribution to stay close to the behavior policy's next-token distribution. RSC works with or without truncated importance sampling (TIS), which caps token importance weights. Our theoretical analysis supports this approach, showing that KL regularization improves convergence guarantees. We evaluate dense and mixture-of-experts models on mathematical reasoning and agentic tasks, including online shopping, tool use, games, and fully asynchronous software engineering. RSC with TIS substantially improves sample efficiency across all agentic benchmarks, attaining higher training rewards with the same number of sampled responses. In final evaluations, it achieves top performance across all six agentic tasks, improving by 14.3% on AppWorld and reaching 9.9 times the strongest baseline's score on BabyAI.

Similarity-Aware Client-Side Reordering for Server-Side Storage Compression in Distributed File Systems

arXiv:2611.22847v1 Announce Type: new Abstract: As growing data volumes strain the capacity of distributed file systems, server-side storage compression transparently reduces disk usage, but its effectiveness is limited when similar blocks land in separate compression records and cannot share matches. We present a client-side reordering framework that transforms the physical block layout before persistence, so that related blocks are co-located within the same compression scope without replacing the existing storage-server compressor. It operates on bounded reordering windows, periodically selects among candidate layouts via lightweight compression of sampled data, and recovers the original logical order during reads with compact window-local metadata. We implement the framework in the SkyFS distributed file system, using the Zstandard (Zstd) compressor on the OpenZFS storage backend. Results on diverse datasets show that the framework reduces physical storage by up to 42.9% relative to logical-order writes, while achieving 94-108% of baseline sequential-read throughput and 91-99% of random-read IOPS. Client-side layout construction is the primary additional write cost, showing that adaptive similarity-aware reordering substantially improves storage efficiency with acceptable overhead.

Sleep2Word: Peripheral-to-EEG Translation via Duration-Aware Wordization

arXiv:2611.21643v1 Announce Type: new Abstract: Estimating full-night electroencephalography (EEG) spectrograms from peripheral measurements could extend EEG-informed sleep analysis to recordings acquired without EEG. This task requires compact representations of signals with different temporal structures. We introduce Sleep2Word, a framework for translating respiration, electrocardiography (ECG), or photoplethysmography (PPG) into EEG spectrograms through duration-aware physiological words. Building on context-dependent chunking, Sleep2Word groups physiological embeddings into variable-duration chunks and jointly quantizes their content and duration into modality-specific wordbooks. A block-autoregressive translator then predicts target EEG word identities and occurrence-specific duration refinements within fixed-duration blocks. This design accommodates independent source and target wordbooks, limits temporal drift, and shortens the autoregressive decoding path. Trained on over 20K PSG cases across 9 datasets, Sleep2Word consistently outperforms existing spectrogram generation baselines in EEG spectrogram reconstruction, sleep staging, arousal detection, and survival prediction. Compressing full night sequences by up to 69.7%, the learned words yield a favorable accuracy-compute trade-off. Sleep event analysis further shows that the learned words capture physiologically plausible associations across modalities, suggesting their potential for interpretable cross-modal analysis.

Temporal Pseudo-RGB Encoding for Label-Free Detection of VEEV Infection in Human Brain Endothelial Cells

arXiv:2611.21892v1 Announce Type: new Abstract: Label-free detection of viral infection is central to antiviral screening and pathogen surveillance. Most deep-learning methods analyze individual frames and do not model temporal progression. We evaluate temporal pseudo-RGB encoding, which stacks three time-offset brightfield frames ($t{-}6$ h, $t{-}3$ h, $t$) as the red, green, and blue channels of a single image so that morphological change becomes color and an ImageNet-pretrained convolutional neural network (CNN) can be trained on infection trajectories without architectural modification. In human brain microvascular endothelial cells (HBMVEC) infected with Venezuelan equine encephalitis virus (VEEV) TC-83, a multi-task network predicts infected-versus-mock status and time post-infection in one forward pass. Across three experimental runs under leave-one-row-out cross-validation, with all 36 wells scored out of fold, the encoding reaches 91.6% accuracy against 87.4% for the single-frame model (paired difference $+4.2$ percentage points, 95% CI $[+2.5, +6.1]$; sign-flip $p = 3\times10^{-5}$; well-level inference, conditional on three runs). At the late-phase endpoint (40-48 h), where floating cells and debris can make conventional brightfield images difficult to interpret, the advantage widens to $+23.9$ points (95% CI $[+19.2, +28.9]$). Time post-infection is recovered on the fixed row split with MAE = 1.31 h ($R^2 = 0.975$). Among the temporal models tested the encoding has the highest mean accuracy, at a fraction of the cost of the recurrent and 3D convolutional alternatives, though only its margin over the single-frame model is resolved; the tested ViT-B/16 backbone showed no corresponding gain, so the finding is limited to convolutional backbones and this virus-cell system.

TIBECG: Temporal Information Bottleneck for Lead-Specific Demographic Privacy in ECG Waveform Release

arXiv:2611.21694v1 Announce Type: new Abstract: Transmitting an electrocardiogram (ECG) waveform can expose demographic information beyond an intended diagnostic task. We evaluate TIBECG, an architectural temporal information bottleneck, on individual leads extracted from hospital twelve-lead recordings. A 2,302,337-parameter processor varies its latent length while retaining the same nine-score consistency objective, without age or sex labels in protector fitting. Across 6,386 recordings, ten folds, 12 leads and 21 lengths, retrospective selections retained at least 95% of each lead's clean macro-F1. Their weaker age/sex interval gain exceeded the best eligible conventional transformation on all leads, by a median 22.08 percentage points. The implemented sampling-rate comparison shows that direct downsampling usually suppresses demographic inference more strongly but at a larger loss of diagnostic utility, whereas TIBECG favors utility preservation. Selection stability also varies by lead. Lower latent Kullback-Leibler (KL) divergence does not reliably rank residual sex leakage at a fixed length. These findings support lead-specific evaluation of waveform processing, not validated consumer deployment or expert-readable reconstruction. When the waveform-input diagnostic classifier is hosted as part of a proprietary cloud service rather than deployed on the device, waveform transmission is required to obtain its scores.

Token Commitment as a Control Surface for Steering Masked Diffusion Generation

arXiv:2611.21669v1 Announce Type: new Abstract: Masked diffusion models repeatedly predict provisional tokens before a sampler decides which predictions become permanent. Existing inference-time steering methods primarily modify activations, logits, or denoising trajectories, while treating this commitment policy as a fixed decoding detail. We instead study token commitment as a diffusion-specific control surface and separate it into two decisions: which tokens to commit and how many tokens to commit at each step. For the first, we introduce counterfactual target-aware commitment, which ranks unresolved positions by a mixture of model confidence and leave-one-token-out target contribution. For the second, we introduce calibrated adaptive budgets driven by target progress and confidence. Across language and vision-language settings in the LLaDA family, multiple targets, two activation-steering methods, and an unsteered control, both mechanisms yield empirically monotonic target-quality trade-offs in every configuration we evaluate. Adaptive schedules also recover competitive operating points relative to scorer-free stochastic schedules matched in average denoising cost. Because both mechanisms require only a scalar score evaluated on a provisional completion, they apply unchanged when the target is an automatic proxy for task success rather than an output attribute: NLI entailment for dialogue summarization and CLIPScore for captioning, neither of which corresponds to a single reusable steering direction. We reproduce the behavior of both mechanisms on a second masked-diffusion family, Dream-7B, for one control target and one task-success target. These results characterize commitment--which provisional predictions become irreversible, and how rapidly--as an interpretable control surface for steering masked diffusion generation under the conditions we test.

TUSI: A Citation-Grounded Graph-Based Benchmark

arXiv:2611.21823v1 Announce Type: new Abstract: Evaluating retrieval-augmented generation (RAG) systems in Persian law remains challenging because benchmarks lack realism, grounding, and coverage of legal reasoning. Existing resources are either synthetically generated from statutory texts, lacking natural information needs, or derived from informal legal question--answer sources without explicit linkage to authoritative legal documents, making them unsuitable for rigorous evaluation of grounded legal reasoning. To address this gap, we introduce TUSI-Bench, a structured Persian legal benchmark designed for the evaluation of RAG-based systems. TUSI-Bench is built upon a multi-layered legal knowledge framework containing 7,261 legal articles and integrating codified statutes, legal commentaries, doctrinal writings, and Supreme Court unification rulings. We also construct a legal knowledge graph linking articles and concepts for structure-aware retrieval. Based on this foundation, TUSI-Bench provides two complementary evaluation tasks: (i) TUSI-QA, an open-ended question answering benchmark containing 1,999 questions that requires grounded responses with explicit legal citations, and (ii) TUSI-Bar, a multiple-choice benchmark containing approximately 500 four-option questions derived from Iranian Bar examination questions. Furthermore, we propose a graph-based retrieval baseline that leverages relational structure within the legal knowledge graph. Experimental results demonstrate that incorporating graph-based relations improves retrieval recall by up to 8 percentage points compared to standard baselines, highlighting the importance of structured legal representations for RAG systems. The proposed benchmark provides a standardized resource for evaluating grounded legal reasoning in Persian and enables future research on structured legal intelligence.