Papers

Modern recommender systems advance not only by scaling data and parameters, but also by encoding task-specific inductive biases through architecture, including sparse feature interactions for click-through rate (CTR) prediction, temporal attention for sequential recommendation, and expert routing for multi-task learning. However, these biases are typically human expert designed or searched within predefined operator spaces. Although Recent LLM-driven code evolution expands this space, unconstrained edits often produce invalid or ineffective architectures, underuse established architecture design knowledge, and fail to preserve successful innovations for reuse. We introduce EvoSkillRec, a promotion-and-reuse framework for cumulative recommender architecture evolution. It first decomposes recommenders into atomic executable skills and represents architectures as typed skill genomes, with each skill equipped with input–output types, semantic annotations, and implementation code. We then evolve models with different tasks through two coupled spaces: a constrained skill–space that mutates, recombines, specializes, and reuses validated skills, and an open-ended code–space in which LLM planners and synthesizers invent new skill modules using prior evolution traces and accumulated experience. An autoresearch controller evaluates candidates, diagnoses failures, retrieves relevant skills, promotes validated innovations into the skill library, and adaptively allocates the proposal budget between the two spaces. Extensive experiments on CTR prediction, multi-task learning, and multi-domain learning, including resource-constrained co-optimization of predictive quality and model FLOPs utilization in generative ranking models, consistently demonstrate the effectiveness of our proposed EvoSkillRec.

Leveraging the strong generative capabilities and stable training dynamics of diffusion models, diffusion-based recommender systems (RSs) have recently emerged as a novel recommendation paradigm, attracting increasing attention from both academia and industry. However, despite the rapid growth of diffusion-based RSs, a critical issue has emerged: the lack of a unified and systematic quantitative evaluation benchmark, which often results in irreproducible experimental results and unfair comparisons across studies due to inconsistent data processing, training configurations, inference procedures, and evaluation protocols. To address this challenge, we propose Eval4DiRec, the first unified and open-source evaluation framework specifically designed for diffusion-based RSs. Eval4DiRec supports 14 representative diffusion-based RS models across five different recommendation scenarios, providing consistent and reproducible experimental settings to systematically assess their performance. Built upon this framework, we conduct extensive empirical studies to benchmark these models under unified protocols. The results highlight the strong potential of diffusion models for recommendation while also revealing key factors and practical challenges that substantially affect their performance, thereby establishing a solid foundation to facilitate fair evaluation and guide future research in this promising field. Our code and data are available at: https://github.com/wangcong2001/Eval4DiRec.

Cross-modal hashing enables efficient retrieval by encoding heterogeneous data into compact binary codes. Recent methods exploit fine-grained relations encoded in multi-label training structure, yet none of them constrains how those relations survive as consistent candidate rankings in finite, multi-length Hamming spaces, which we term the relevance resolution bottleneck (RRB). To address the RRB, we propose MultiBit, which transfers relevance resolution from multi-label structure to multi-length Hamming spaces. MultiBit first constructs a scale-decomposable fractional relation teacher from dataset-level label co-occurrence and label specificity, and models dependencies from local to long-range over continuous diffusion scales. It then maps the discretized diffusion scales and their quadrature weights to scale-aware bit subblocks of the maximum-length code, organizes the target code lengths as nested prefixes, and aligns their Hamming candidate rankings with the teacher relations. Experiments on multiple benchmarks demonstrate improved retrieval accuracy. Code is available in the supplementary material.

Memory is critical for AI agents. Many existing agent-memory systems follow an Ahead-of-Time (AOT) design, constructing memory before a specific request arrives. While this reduces online serving cost, such request-agnostic memory construction can discard fine-grained information that later becomes important. To address this limitation, we propose Just-In-Time Agent Memory (JAM), a trainable framework for query-conditioned context construction at runtime. A Memorizer preserves complete raw histories in a hierarchical page-store with compact navigational summaries, while a Researcher iteratively retrieves, inspects, and integrates evidence for each request. To train these memory-use behaviors, we introduce Memory-Gym, an evidence-grounded data synthesis pipeline covering nine task types across six domains, and optimize the Researcher through verified-trajectory supervised fine-tuning followed by Hint-guided Group Relative Policy Optimization. We demonstrate the effectiveness of JAM across a variety of benchmarks on agent memory and long-context processing, where it achieves stronger task performance than AOT-style memory systems while remaining substantially more efficient than prior trained agentic memory approaches. To support reproducibility and future research, we release our anonymized source code at https://github.com/VectorSpaceLab/general-agentic-memory.

Product attribute value extraction (AVE) is a fundamental task in e-commerce, aiming to identify specific values of predefined attributes from multimodal product profiles such as text and images. While multimodal large language models (MLLMs) have shown promise for AVE, they face challenges in extracting implicit attributes that require joint reasoning over visual and textual cues, often confusing semantically similar values. However, existing methods often fail to resolve such ambiguities because the correct value often depends on subtle multimodal cues that are easy to miss or override. To address this challenge, we propose Correcting to Predict (C2P), a framework that treats attribute extraction as a correction process. Given an initial pseudo-value such as a retrieved candidate or placeholder, the model learns to correct it using multimodal evidence. During training, diverse pseudo-values help the model learn evidence-based correction behavior, and a self-consistency refinement stage further reduces sensitivity to pseudo-value perturbations. At inference, a fixed placeholder triggers the learned correction behavior, enabling efficient single-pass prediction without online retrieval or iterative refinement. We evaluate C2P on a public benchmark and a large-scale industrial dataset. Offline results show that C2P outperforms strong baselines, with notable gains on ambiguous attributes. Online A/B tests on AliExpress further show consistent improvements in seller adoption, attribute completeness, and user engagement, validating C2P's effectiveness and efficiency in real-world deployment.

We describe our system for DocSem, the document-grounded quantitative reasoning shared task at DocInsights 2026, and analyze why it succeeded on labeled data and failed on the test set. The pipeline pairs hybrid block retrieval with Program-of-Thoughts (PoT) generation executed in a sandboxed interpreter, self-consistency sampling, and entity enrichment from chunk-level knowledge graphs. On our held-out split, application architecture moved the metrics far more than model scale did: PoT added 0.282 joint accuracy to a compact 7B model but at most 0.005 to a 72B model, and a 27B model with the full harness matched the 72B (0.884 vs.\ 0.873) at roughly 2.7$\times$ fewer parameters and a quarter of the CO$_2$. We read this through a distinction between world knowledge, which scales steeply with parameters, and language knowledge, which scales gently, and show that structured-output training makes a compact model harness-ready rather than merely small. On the raster, watermarked test PDFs the same system collapsed to 13.58% joint (rank 149 of 163); a controlled re-rendering of the validation set reproduces the OCR half of the collapse while bounding what the simulation misses. Auditing the physical nature of evaluation inputs precedes architecture, and the leaderboard's bimodality is consistent with reading quality, not reasoning, having separated the field.

Semantic ID (SID) based generative recommendation represents each item as a sequence of discrete tokens, and recommends by generating the SID of the item a user would like to interact with. Both dominant paradigms in this domain generally pay for generation token by token: autoregressive models decode the tokens left-to-right, while non-autoregressive models decode in parallel yet still need multiple rounds of refinement to stay competitive. Therefore, both generally spend multiple forward passes per item, a cost that is prohibitive in latency-sensitive recommender systems. We ask whether an item can be generated in a single forward pass, and answer it through a new perspective which we call average probability velocity. We view SID generation as a flow of token generation probabilities and characterize it by its average velocity over the whole generation process. We prove that this average velocity is fully determined by the average generation probability of each token. Therefore, we directly parameterize and learn the probabilities of all tokens in a single forward pass with a bidirectional Transformer. As these probabilities are generated independently across positions and the coherence among tokens is lost, we further design a dual-level flow contrastive objective to restore the coherence among an item's tokens. It contrasts the target SID against negative SIDs at both the token and SID levels. The token level ranks the generation probabilities of the target tokens above those of negative SIDs, while the SID level scores the tokens of each SID as a whole item for capturing token coherence of each item. Extensive experiments show that our model not only generates recommendations far more efficiently ($8.39-10.04\times$ speedup over the second-fastest AR/NAR method) but also attains superior recommendation accuracy ($7.77\%$ average improvement over the second-best.

In large language model-based recommender systems, identity cues embedded in prompts can steer recommendations toward group-level patterns even when the underlying behavioral evidence remains unchanged. We introduce PromptShift, an interpretable, training-free framework for quantifying and mitigating such identity-cue preference drift. We define Drift as the divergence, in both item membership and ranking order, between a recommendation list generated under an identity-cued prompt and the reference list produced from the same user's interaction history alone. SliceShift then measures the extent to which a cued list gravitates, relative to the history-only reference, toward items that are more popular within the cued slice than among the global user population. Beyond conventional accuracy, we propose DifHitRate, a difficulty-weighted hit metric that credits only relevant items, assigning higher credit to hits that are less popular within the cued slice and ranked higher in the list. All components are supported by an identity-slice-by-item table constructed from positive interactions, which further enables an adaptive post-hoc reranking strategy: the reranker interpolates between the original LLM ranking and inverse slice-popularity, with personalized interpolation weight. Experiments on two datasets with three LLMs show that identity-cued prompts incur higher mean Drift than identity-free paraphrase controls, an effect beyond generic wording sensitivity, and that SliceShift is positive across all six dataset-model settings. PromptShift consistently reduces both Drift and SliceShift, lowering macro-mean SliceShift by 62.42%, while improving DifHitRate, HitRate and MRR. These results demonstrate that identity-cue preference drift can be measured and mitigated without any model training, albeit with a modest, metric-dependent utility cost.

Does conversational memory need LLM-extracted facts, or is selecting the right raw turns enough? Published results disagree. Extraction-based systems report gains from distilled facts. Recent studies find raw history with good ranking does as well, but disagree about whether ranking matters. We ran a pre-registered study on held-out LoCoMo conversations and LongMemEval. At a tight budget on LoCoMo, raw turns selected by a single call to Jev, a typed decision model, are non-inferior to an LLM-extraction memory (one-sided 95% bound -3.0 points against a -5-point margin). Blind human grading narrows the margin but does not change the result. Raw turns cost 3,061 times less to write, and the result holds with a second answer model. Within this study, reranking's gain shrinks as the budget grows. It adds 17.4 points on LoCoMo and 9.1 on LongMemEval when three of 30 candidates are kept. At generous budgets it adds 1.5 and 1.1, and extraction systems are more accurate. This suggests why published results disagree. At matched context, Jev selects as accurately as an LLM reranker (non-inferiority bound -2.0) at a third of the latency, and more accurately than a multi-call graph traversal. Reranking lowers correct abstention. Plans, code and graded answers are released.

Multimodal language models rerank visual document retrieval results accurately, but scoring every candidate page at full cost makes them slow. Some methods that compress these rerankers need relevance labels to regain accuracy, and they rank by the reranker score alone. RidgeRank measures how much relevance signal the reranker score lacks and recovers it from the retriever score through a closed-form fusion rule. Maximizing a correlation objective gives the optimal fusion weight, along with the exact condition under which the reranker score by itself cannot reach that optimum. The reranker is further corrected by a single vector applied to an intermediate hidden state, obtained through one centered ridge regression onto the same model's full-depth scores on uncompressed pages. On 12 datasets drawn from ViDoRe 2 and ViDoRe 3, evaluated with two retrievers and two language model backbones, RidgeRank brings NDCG@5 to within 1.2 pp of a full cross encoder with speedups of up to 48 times, advancing the accuracy and latency Pareto frontier for visual document reranking.

Multi-hop retrieval-augmented generation requires a retriever to connect evidence distributed across documents while preserving a concise, faithful generation context. Existing indexes leave two complementary gaps: chunk-based RAG can break cross-passage evidence chains, whereas an unlabeled pairwise projection without generating-topic provenance cannot jointly preserve topic-level co-participation and per-occurrence entity descriptions. We propose STITCH-RAG, a hypergraph-based framework with three coupled components. First, a semi-merged topic hypergraph encodes multi-entity co-participation as topic-summary hyperedges while retaining per-chunk entity states linked by canonical-name equivalence. Second, spatio-temporal influence bridging propagation (STIBP) combines topic-space propagation with deterministic chunk-index linkage across name-equivalent states under frequency-adaptive decay. Third, continuous STIBP scores replace binary entity-match seeds in localized Personalized PageRank (PPR). We characterize the condition under which this prior assigns more PPR mass to ground-truth evidence than a binary prior. Under the reported protocol, STITCH-RAG attains the highest reported Contain-Acc and LLM-Acc point estimates among the compared methods on HotpotQA and 2WikiMultiHopQA, and higher Recall@8 than the methods included in the standardized retrieval comparison. Results on the mixed-domain benchmark remain auxiliary preference-based evidence because only LLM-judged accuracy is available.

Retrieval-augmented generation (RAG) grounds large language models in external sources, but retrieved passages often name the right entities without providing the facts needed to answer. Even when instructed to abstain, 12 generators answer 40.0-99.3% of insufficient-evidence questions. Training generators to abstain ties the decision to model weights, may reward answers recalled from parametric knowledge, and still requires a full generator call. Can sufficiency be judged from the question and evidence alone, before any answer exists? We identify pitfalls in constructing insufficient-evidence tests: removing relevant evidence or pairing evidence with unrelated questions can reveal labels through lexical overlap or evidence position. We build a paired benchmark using substitution, deletion, and question-swap constructions that vary answer support while controlling selected surface features, such as word use. Sufficiency can be judged without generating an answer, but no single signal works across all datasets. We introduce RINSE (Relevance Is Not Sufficient Evidence), which combines three signals: whether every part of the question is covered, whether any passage offers an answer, and whether a small language model reading the passages together judges them sufficient. Across six datasets, RINSE ranks sufficient above insufficient evidence with a score of 0.837 (chance 0.5), exceeding the best of 10 prior methods (0.746) and a frontier model queried through an API (0.784). Its weakest dataset scores higher than any other method's weakest (0.684 vs. 0.676). RINSE runs locally before generation, taking 36.5 ms per question on a single GPU.

Semantic Textual Similarity (STS) methods assume that a document's lexical content faithfully represents what it asserts. This assumption fails for discursive documents that discuss, compare, critique, and contextualize other positions in the process of articulating their own. The result is semantic diffusion: similarity scores between documents are inflated by vocabulary acquired through discursive engagement rather than substantive alignment. Standard Natural Language Processing (NLP) preprocessing (tokenization, stopword removal, stemming, lemmatization) cannot address this problem because it operates at the lexical level, treating all content identically regardless of its discursive function. This paper introduces high-level text preprocessing: a systematic, rule-based intervention applied before the standard preprocessing pipeline to isolate each document's actual claim from its discursive structure. We propose 12 rules, each with an explicit rationale, and demonstrate their effect on an encyclopedic philosophical corpus: three entries from the Stanford Encyclopedia of Philosophy (virtue ethics, deontological ethics, and consequentialism). A three-phase experiment using eight Transformer-based STS models shows that preprocessing reduces centroid cosine similarity scores across all three theory pairs, with 23 of 24 model-pair comparisons showing the expected decrease and cross-model agreement ranging from 7-1 to 8-0. We introduce the semantic diffusion index (SDI), a per-document metric for assessing the semantic reorientation between a document's raw and high-level preprocessed representations. Although the framework is demonstrated using philosophical texts, it potentially addresses a domain-agnostic problem applicable to legal texts, policy documents, academic articles, and any genre in which a discursive approach introduces vocabulary from positions the document does not endorse.

Graph Neural Networks (GNNs) perform well on homophilic graphs but struggle in heterophilic settings, where connected nodes often carry dissimilar labels. Existing evaluations typically compare architectures under a fixed node-feature representation, leaving unclear whether conclusions about heterophily robustness remain stable as the input representation changes. We address this question by constructing parallel feature variants of two large-scale heterophilic benchmarks, Roman-Empire and Amazon-Ratings, pairing each graph with representations ranging from static fastText vectors to contextual Transformer embeddings and evaluating seven GNN architectures across these representations. We find that the effect of representation varies across architectures: on Roman-Empire, the contextual gain ranges from 2.38 percentage points for GCN-sep to 13.67 points for GAT, with H2GCN gaining 8.77 points. On Amazon-Ratings, where node text is limited to short product titles, GAT improves by 6.78 points from fastText to MPNet, while GCN-sep changes by only 0.20 points. These results show that architectural performance is conditional on node representation: the same representation change can produce different magnitudes of performance gain across architectures, so architecture and representation cannot be treated as independent evaluation factors. A rank-correlation analysis on these two benchmarks further shows that the relative ordering of architectures remains highly stable across representations, isolating differential sensitivity, rather than ranking instability, as the primary effect.

Signal-multiplexing network ranging (SM-NR) shares broadcasts across node pairs, but its sequential operation leads to a ranging cycle that grows linearly with network size. This paper proposes a concurrent coded SM-NR (CC-SM-NR) framework for asynchronous half-duplex networks. Firstly, the CC-SM-NR protocol coordinates concurrent transmissions through binary transmit-listen codewords. The transmit-listen schedule defined by these codewords ensures reciprocal observations subject to a finite concurrency limit. Then, we derive the exact minimum number of transmit-listen rounds without a concurrency limit, which reveals that the minimum grows logarithmically with network size. To account for practical scenarios, we establish the necessary and sufficient conditions for the constant-weight feasibility of codewords under a finite concurrency limit. Subsequently, we propose a low-complexity scheduling algorithm that achieves the minimum round count within the constant-weight codeword class. To support higher observation redundancy, this scheduling design is extended through a greedy construction. Finally, simulation results demonstrate the effectiveness of the proposed schemes for network ranging.

Generative recommenders retrieve items by generating identifiers, but a valid identifier can remain outside the beam after catalog expansion. This raises two connected questions: which failures can identifier assignment repair, and how should retrieval proceed beyond the initial beam? We characterize assignment repair with a fixed generator and retained old identifiers. Output-invariance certificates identify failures shared by all admissible assignments. Under a common effective prefix, coupled support and ranking constraints give the exact feasible interval of new-item counts for target recovery. Building on this characterization, Beyond the Beam (BB) obtains minimum-replacement repairs through an integral flow formulation, selects a shared map and adapts the generator. At inference, generative likelihood and collaborative evidence define one score for ranking, candidate priority and stopping. Retained prefix bounds guide candidate completion and certify its global Top-$K$ when the stopping condition is met. Exhaustive finite-catalog evaluation confirms construction in every feasible case. Across three Amazon Reviews categories and three random seeds, the full T5 procedure improves mean Recall@10 by 15.5–46.3% and NDCG@10 by 15.2–44.4% over the best-performing evaluated generative baseline for each dataset and metric. Matched controls show that shared construction and adaptation improve new-target ranking and certification efficiency on Beauty and Toys. Combined scoring and candidate completion improve NDCG@10 across all three datasets with both T5 and decoder-only LC-Rec.

Multimodal large language models can expose task-relevant evidence through generation, but producing useful evidence does not by itself determine how it enters a retrieval embedding. We study whether the semantic organization of that evidence can also specify where representations are read. To address this question, we introduce EviAlign, which couples Semantic Evidence Generation with Boundary Readout in a shared multimodal large language model. It organizes evidence into five semantic units, reads the contextualized state at each unit boundary, and aggregates these states into a single normalized embedding. Generation and contrastive retrieval objectives jointly train this shared structure. With the same trailing readout, semantic evidence and free-form CoT yield nearly identical retrieval performance, suggesting that evidence organization alone does not explain the full gain. A controlled $2\times3$ study compares consistent and permuted evidence organization across three readout strategies, using training targets with matched evidence spans. With five readout states and the same mean pooling, the advantage of consistent semantic organization grows from 0.65 points at length-based training positions to 2.39 at evidence boundaries, yielding a 1.74-point co-design interaction. Across 12 MMEB retrieval tasks, EviAlign achieves 76.9 average Recall@1 with 500K training pairs while retaining single-vector indexing and scoring.

Guideline documents are published as unstructured PDFs whose evidence is locked in visual structures—tables, flowcharts, and graded recommendations—that standard retrieval pipelines flatten into fixed-size text chunks. We cast evidence access as a document image analysis problem: parse each page image into typed structural elements, then retrieve structure-aware evidence units that follow the document's own layout (sections, table rows, flowchart paths, graded recommendations), each keeping its structural context so a result points to a specific element rather than a page. On 26 clinical practice guidelines from 9 sources (3,619 pages, Chinese and English) with 199 evidence queries, structure-aware units rank the gold element first under BM25, dense, and hybrid retrieval (hybrid Element Hit@1 of 0.382), with a significant element-level ranking gain over per-element OCR text (MRR_e +0.107, p=0.002; the Hit@5 gain is directional, p=0.17), while matching page-level recall (Page Hit@5 0.879 vs. 0.889, p=0.75) at 3.8x less context and clearly outperforming a ColPali visual-RAG baseline (PH@5 0.497).

Sign-aware recommender systems have recently been developed to leverage negative feedback for a deeper understanding of user preferences. However, our empirical diagnosis reveals that state-of-the-art graph-based sign-aware recommender systems are paradoxically valence-blind. Even though they explicitly incorporate sign information during training, they consistently fail to differentiate liked items from disliked ones at the ranking stage, frequently infiltrating top-K recommendations with disliked content. Through linear probing, we show that while valence information exists in the learned embeddings, it remains inaccessible to the inner-product scoring function. This widespread failure remains entirely undetected because conventional evaluation metrics, such as Recall, HR, and NDCG, assign a uniform utility of zero to both negative and unobserved items, creating a systematic evaluation blind spot. To bridge this gap, we propose a family of signed metrics, Signed Recall, Signed HR, and Signed NDCG, that explicitly penalize the recommendation of disliked content. Systematic re-evaluation under our proposed metrics fundamentally reshapes the established performance landscape, revealing that methods ranked highly under conventional metrics often fail to protect users from disliked content. Finally, through a proof-of-concept auxiliary loss, we confirm that the proposed metrics provide actionable training signals, guiding models toward valence-aware behavior without sacrificing conventional relevance. For transparency, our source code is available at: https://anonymous.4open.science/r/signed-rec-benchmark-07E4

Large Language Models (LLMs) enable us to better understand text documents, including PDFs and Word documents. However, LLMs, as well as more modern LLM agents, i.e., those with tool-calling abilities, typically treat such documents as plain text, ignoring the fact that they are often organized hierarchically into sections and subsections. Extracting this structure, while difficult, can improve efficiency and effectiveness for agents (and humans)—since only sections relevant to a given task need to be processed. Unfortunately, prior work on structure extraction provides no formal guarantees on how well the inferred structure matches the true one. Instead, we target a robust and compact variant that is feasible to infer and useful in practice. Robustness ensures that the text under each subsection header is a superset of the text under the same header in the true structure. Compactness seeks to minimize this superset, reducing agentic cost (or human cognitive load). We propose SHED, a two-stage workflow for inferring a robust and compact structure. The first stage is pluggable with an infinite family of approaches, each guaranteeing robustness for a specific document class. We theoretically characterize the document space using these classes and their hierarchical relationships. Empirically, SHED improves F-1 scores (measuring the robustness–compactness trade-off) by 13%–68% over non-LLM baselines and 9%–15% over expensive LLM-based approaches. Finally, we show how SHED-inferred structures are valuable for agentic document analysis: agents using SHED outperform baselines, achieving 3%–23% higher accuracy while being up to 10x cheaper.

Scientific literature search often begins with an open research problem rather than a known target paper or a fixed candidate set. We introduce INSPIRE, a benchmark for evaluating agents that search prior literature to make progress on solution-redacted research problems. Each instance pairs a research brief with a target-specific cutoff three months before a later paper and evaluates ranked outputs against graded cited antecedents from that paper's realized research lineage. Search proceeds over an open corpus, while the identity of the target paper and membership of its cited antecedents remain hidden from the agent. Beyond end-to-end retrieval quality, INSPIRE uses logged search trajectories to distinguish three coupled stages: resource exposure, whether useful antecedents are surfaced during search; selection, whether exposed antecedents are retained; and ranking, how effectively retained papers are ordered. Across 476 computer-science targets under a shared search interface and budget, the strongest evaluated agent achieves 0.284 nDCG@10. Results show that current agents more readily recover an isolated antecedent than assemble a broader portfolio of relevant prior work. The stagewise analysis identifies resource exposure as the largest observed bottleneck, with further losses in selection and ranking. We additionally construct replay-valid hindsight demonstrations and show that they improve held-out search without changing test-time information, establishing that the benchmark provides an actionable learning signal. INSPIRE therefore enables both end-to-end comparison and stage-resolved diagnosis in a setting where the agent must construct its own working criterion of relevance.

In this work, we consider algorithmic harms that may arise as generative models are incorporated into machine learning platforms. We argue that existing harm taxonomies and threat models require extension to (1) address novel causal drivers of well-studied representational and quality-of-service harms; and (2) anticipate and mitigate endogenous harms, such as sanitization, which may arise when system inputs are misaligned with the system designer's objectives, or the generative model's inductive priors. To this end, we introduce an expanded taxonomy of algorithmic harms associated with the use of generative models in non-conversational recommendation systems. In addition, we offer a causal analysis of how problematic subsets of the (input, output) joint distribution can arise, in an effort to inform harms detection and mitigation efforts.

The RecSys Challenge 2026 studies conversational music recommendation as a joint item recommendation and response generation problem: given a multi-turn dialogue, systems must retrieve relevant tracks from a large catalog and produce a grounded natural-language response. This paper presents the challenge task, dataset, evaluation protocol, and official results. Beyond the leaderboard, we analyze the 16 accepted systems through a common retrieve–rerank–generate framework and examine how recommendation performance varies across users, requests, and dialogue contexts. Strong systems commonly combine heterogeneous candidate sources and preserve source-specific evidence for learned reranking. Across the system papers and our organizer-side analysis, robust design also means 1) grounding cold-start retrieval in multi-turn conversation and item signals, 2) using intent detectors, and 3) modeling the full multi-turn context rather than the current query alone. We further identify limitations of the benchmark and evaluation protocol, including single-ground-truth relevance and teacher-forced evaluation of synthetic dialogues. Together, these findings provide practical guidance for future conversational recommender systems and shared evaluation efforts.

Evaluating recommender systems before deployment requires simulators that remain realistic over long, closed-loop interaction cycles. Current LLM-based agents generally function as surface imitators that generate realistic short-term responses but fail to capture the underlying user dynamics such as fatigue, affect, and shifting preferences. We propose PsyBer-Agent, a psychology-inspired simulator that models these dynamics through its Psy-Engine. This engine represents each user with three evolving, model-defined latent states. To calibrate simulated behavior against real interaction logs, we align log-conditioned reference trajectories with simulator-generated latent-state trajectories via Unbalanced Gromov–Wasserstein optimal transport. This approach provides structured supervision for a lightweight policy without the computational burden of fine-tuning the backbone LLM. We also introduce WebSim and the PsyBer Benchmark, which encompass multimodal scenarios and seven recommender backbones. Our framework evaluates behavioral fidelity, latent-state consistency, robustness, and responses to social-cue perturbations. On the tested horizons, PsyBer-Agent shows lower behavioral discrepancy than the prompt-only simulators while maintaining stable performance under large-scale concurrent simulation for up to 5M queued simulation tasks under bounded concurrency. Code is available in the supplementary material and at: PsyBer-Agent: https://anonymous.4open.science/r/CC66F/, WebSim Platform: https://anonymous.4open.science/r/D8EAX/.

Agentic retrieval-augmented generation (RAG) interleaves reasoning with repeated retrieval, giving the retriever influence over both the evidence an agent observes and its subsequent search decisions. We study retriever backdoors that exploit this feedback loop and repurpose weak backdoor purification to conceal their presence. An attacker supplies a compromised retriever checkpoint while leaving the search agent and deployment corpus unchanged. Without corpus write access, the attacker can still suppress useful evidence, persistently retrieve a selected existing document, or steer the agent toward prolonged search, inflating retrieval, context, and latency cost. To conceal these behaviors from detection, we propose leveraging a controlled inject-and-remove cycle: deliberately inject a weaker backdoor and then unlearn it. This process weakens detector-visible signatures and fools the backdoor detectors with an illusion of purification while preserving the malicious retrieval behavior. These findings expose a systematic vulnerability in RAG systems in which a weak defense becomes an attacker's concealment tool for a backdoored retriever, even when the underlying corpus remains trustworthy.

Open-domain question answering (ODQA) over multimodal document corpora requires linking evidence scattered across text, tables, and figures. Existing systems often store these sources separately or retrieve only coarse pages, which weakens global evidence linking. We present PILAR, a page-grounded unified evidence representation instantiated as an entity-linked assertion graph. PILAR maps sentence-, table-, and figure-derived facts into a common assertion space and uses the graph as a controlled linking layer over robust page retrieval. In a shared-reader evaluation with four agent frameworks, fourteen retrieval backends, and two benchmarks, PILAR achieves the best end-to-end EM/ANLS. Gains are largest on compositional, cross-document, and multimodal questions, with a single-shot improvement of +1.6 EM over flat retrieval, rising to +2.9 on compositional and +5.9 on 3-hop questions. Ablations show that current gains are driven mainly by the text-instantiated slice of the framework, while visual assertions help only after locality-aware filtering. We therefore position PILAR as a unified evidence representation for multimodal ODQA rather than a standalone visual-reasoning module.

Recommender systems rely heavily on heterogeneous behavioral feedback to infer user preference. Although abundant, these signals are imperfect measurements: the same observed behavior can arise from different underlying states, such as genuine enjoyment, passive consumption, or inattention. The challenge is especially acute in short-form video, where watch-based signals are strongly affected by measurement confounders such as video duration - the same watch time can imply different levels of preference for videos of different lengths, while ratio-based metrics can systematically favor short videos. As a result, optimizing raw engagement can amplify measurement artifacts rather than improving user value. We propose a Factorized Latent Value Model (FLVM) for measuring user preference from heterogeneous behavioral feedback. The model treats observed behaviors as noisy measurements of a low-dimensional, factorized latent value state and uses structured output heads to model heterogeneous feedback signals. A restricted baseline path captures predictable variation from measurement-confounding features such as video duration, user propensity, and session context, while a routed latent path estimates preference-relevant value advantage. The resulting latent value score can be integrated into an existing recommender system as a ranking feature or ranking score. On YouTube Shorts, a major short-form video platform, this model improves offline metrics and lifts a primary viewer enjoyment metric by 2.67% in online A/B tests.

Visual document retrieval requires expressive representations to match queries with evidence distributed across text, tables, and page layouts. Multi-vector representations capture fine-grained information, but storing and comparing many vectors introduces substantial retrieval costs. In this paper, we introduce RandSlot, a simple approach to learn compact visual document representations with random soft tokens. During training, we append independently-sampled random unit vectors to query and document input sequences and resample them at every use, without introducing learnable soft-token parameters. The encoder contextualizes these auxiliary inputs with the original content to produce a small set of retrieval vectors. A standard late-interaction objective trains the encoder to extract relevant information under varying input conditions. Experiments with different backbone models show that RandSlot improves retrieval quality over alternative readout strategies under the same vector budget. Further analysis shows that these gains can persist when random soft tokens are replaced with zeros at inference, demonstrating that random inputs during training can improve compact retrieval representations even when inference no longer requires sampling.

Cross-modal retrieval has been advanced by vision-language pre-trained models that encode images and texts into a shared dense embedding space. While dense representations effectively capture overall semantic similarity, they often obscure fine-grained visual-textual information needed for precise cross-modal matching. Recent methods introduce a learned sparse branch to complement dense matching with lexical evidence, but they rely on a redundant language-model token space and lack explicit grounding for sparse dimensions. We propose GRASP, a compact and grounded sparse learning framework that mines visual-textual concepts from the corpus. A lightweight sparse head is trained to predict concepts relevant to each image or text, yielding interpretable concept-level evidence that complements dense semantic matching. Extensive experiments show that GRASP improves retrieval accuracy over the state-of-the-art dense-sparse baselines while yielding a more compact and grounded sparse space.

Work on memory for multimodal agents optimizes what is written, updated and retrieved. Between retrieval and the answer, however, is a stage that multimodal memory evaluations do not isolate: what of the retrieved memory reaches the model, and in what form. We call it delivery, and a controlled decomposition on MemLens locates the remaining room there. With the retrieved evidence set exactly fixed, delivering the original pixels instead of withholding them raises accuracy by 13.87 points on an 8B backbone, whereas making retrieval perfect on those same messages improves it by 2.31. Delivery is the larger term on all three MemLens backbones and grows with backbone strength; retrieval grows too, without closing the gap. We propose DeliverMem, an instantiation of delivery as three decisions: keep the original modality, give each item a readable identity, and state when it was seen, with a retrieval-side adapter for the one property delivery cannot supply. Each is measured against a delivery-matched control that alters only its own variable. DeliverMem leads the strongest published memory agent on MemLens at all four context lengths, and beats DMV-Bench's own strongest method at every setting on both backbones. On MemLens it does this on a tenth to a seventieth of the input. Each decision helps only where the question lacks what it supplies, and is null elsewhere. A single fixed configuration nonetheless leads both benchmarks, without training any component or modifying the stored records. Project page: https://avalon-s.github.io/DeliverMem/