Computation and Language · Cross-listed from cs.LG
ARMED: Adaptive Retrieval-Augmented Medication Recommendation
Kyungkeon Chung, Se Un Park, Hyejin Lee et al. (+1 more)
Medication recommendation requires jointly considering diagnoses, comorbidities, longitudinal patient history, and treatment interactions. Although recent
retrieval-augmented generation approaches using
language models have shown promise for clinical recommendation, most existing systems rely on fixed
retrieval pipelines that cannot adapt to diverse clinical
retrieval scenarios and may introduce noisy or irrelevant evidence. We propose ARMED, an adaptive clinical RAG framework that dynamically expands
retrieval according to
retrieval sufficiency and current-admission relevance. ARMED first observes the initial
retrieval outcome and selectively performs supplementary or fallback multi-query
retrieval when additional evidence is needed. The framework also incorporates a relevance gate that activates only clinically relevant same-patient history and organizes retrieved evidence into structured query-level clinical evidence. Experiments on MIMIC-IV demonstrate that ARMED consistently outperforms previous static
retrieval baselines and improves overall medication recommendation quality. By combining locally runnable small
language models with adaptive
retrieval, ARMED supports medication recommendation without relying on external
language model APIs or task-specific fine-tuning.
Read the full abstract ▾
All authors Kyungkeon Chung, Se Un Park, Hyejin Lee, Junyoung Park
Computation and Language
CultureBot: A Framework for Retrieval-Augmented Exploration of Cultural Heritage Data
Vasileios Karampinis, Angelos Vlachos, Giorgos Filandrianos et al. (+4 more)
Digital cultural heritage collections encode rich relations among objects, places, periods, providers, images, and source institutions, yet users typically access them through keyword search and faceted browsing. We present CULTUREBOT, an interactive framework for natural-
language exploration of cultural heritage collections that combines knowledge-graph-guided
retrieval, lexical-semantic
retrieval, and source-linked generation. CULTUREBOT maps user queries to structured catalogue relations, enriches retrieved records with textual evidence, and generates answers that link back to the underlying repository records. We instantiate the framework on 81,852 SearchCulture records and release an online demonstration. In an expert evaluation involving six cultural-heritage specialists, 55 persona-grounded queries, and 250 evaluations, the graph-guided system was rated significantly higher than two general-purpose assistants and a non-graph ablation, with the largest gains in evidence grounding, diversity,
reasoning, and completeness. These results support graph-guided
retrieval as a practical way to expose the relational structure of cultural collections through conversational access while preserving links to catalogue evidence.
Read the full abstract ▾
All authors Vasileios Karampinis, Angelos Vlachos, Giorgos Filandrianos, Maria Lymperaiou, Paraskevi Tzouveli, Chrysoula Zerva, Giorgos Stamou
Computation and Language · Cross-listed from cs.AI
RetroPlanner: Multi-Route Expansion and Multi-Objective Reasoning for LLM-Based Retrosynthesis Planning
Ye Rin Kim, Yongjun Jeong, Seongjun Yun et al. (+4 more)
Multi-step retrosynthesis planning is often considered solved once a route to purchasable starting materials is found, but such a route should also be practical in real-world chemistry. Recent LLM-based planners formulate this process as sequential decision-making over an AND-OR graph, yet typically expand a single molecule per turn and are optimized for solvability alone, leaving alternative routes uncompared across the multiple objectives that determine practicality before the search budget is committed. We introduce RetroPlanner, an LLM-based planner that jointly compares multiple routes to steer the search toward higher synthesizability, as measured by reaction plausibility, round-trip consistency, and starting-material price. RetroPlanner is trained to reason over chemical evidence: structural progress from the target molecule, bond changes and reaction classes, where candidate routes diverge from one another, and how they rank across multiple objectives. Experiments on multi-step retrosynthesis
benchmarks demonstrate that RetroPlanner recovers more alternative routes than prior LLM-based planners while improving synthetic route quality and maintaining a high success rate.
Read the full abstract ▾
All authors Ye Rin Kim, Yongjun Jeong, Seongjun Yun, Hanbum Ko, Soohyun Oh, Sungbin Lim, Sungwoong Kim
Computation and Language
PersonaFactor: Preference Factors with Dynamic Selection for LLM Personalization
Hang Su, Zequn Liu, Jianghan Chao et al. (+3 more)
Although users may express different personalization needs across tasks, these task-specific behaviors can reflect shared underlying preference mechanisms. Existing LLM personalization methods often focus on surface-level interactions or undifferentiated histories, overlooking the shared mechanisms underlying these behaviors. We introduce PersonaFactors, a framework that distills reusable and interpretable personalization factors from heterogeneous user interactions. We construct a library of thirteen PersonaFactors by extracting evidence-grounded rationales and clustering recurring patterns across tasks, and develop a dynamic factor-guided generation framework that selects factors relevant to each user's history and current request. Across four personalization
benchmarks and four backbone models, Dynamic PersonaFactor achieves the best score on every reported metric. Representation analyses further show that PersonaFactors and user history induce partially overlapping low-dimensional representations, and that their joint use yields complementary shifts. Taken together, these findings suggest that factor-based and history-based personalization can mutually reinforce each other.
Read the full abstract ▾
All authors Hang Su, Zequn Liu, Jianghan Chao, Qiuyu Tian, Xuesong Lu, Yingce Xia
Computation and Language
Rule vs.Consequence: Dissociable Internal Representations of Moral Reasoning in LLMs
Eugenie Shi
Large
language models are increasingly deployed in morally and legally sensitive contexts, yet little is known about how they internally represent distinctions between moral
reasoning styles. We ask whether a model's differential response to rule-framed and consequence-framed moral questions reflects distinct internal structure or merely surface lexical sensitivity. We construct a 2$\times$2 factorial
dataset crossing
reasoning style (rule-based vs. consequence-based framing) with judgment polarity (permissive vs. prohibitive), drawn from three established moral
reasoning benchmarks and evaluated on a cleanroom test set of novel phrasings that includes conflict pairs where the two framings predict opposite labels. Using linear probing and activation steering, we identify two orthogonal directions in the residual stream: Dir-A tracks
reasoning style and Dir-B tracks judgment polarity, each near-blind to the other factor (cross-correlation $|r| < 0.04$). This double dissociation replicates across three instruction-tuned models from different developers at comparable proportional depth, and it is not knife-edged in layer: the directions are recoverable and causally effective throughout an extended band of the network. Dir-A can causally change judgments, and in conflicting scenarios, it selectively corrects errors in rule-based judgments without changing consequence-based judgments, suggesting that it does not simply make the model more likely to accept everything. Generalization beyond the training templates is partial: the direction transfers across
benchmark sources, but its advantage over lexical baselines on novel phrasings is modest. We therefore frame the result as moral
reasoning style sensitivity, a geometrically dissociable and causally effective structure, rather than internalized normative
reasoning.
Read the full abstract ▾
Computation and Language
HexaBee-4B: Six-Language Translation with Multilingual Answering and Bidirectional Supervision
Iman Barati, Mohammadmostafa Rostamkhani and Mohammad Badzohreh
Multilingual translation requires a model to preserve meaning while producing text in a requested
language. We present HexaBee-4B, a translator adapted from Gemma 3 4B IT through three stages: Wikipedia continued pretraining, joint multilingual answering and answer-text translation, and bidirectional translation specialization. The intermediate stage connects two tasks through shared multilingual answers: answering a question in the requested
language and translating an answer into that
language. It comprises 19.46 million answering records, including 5.76 million cross-
language question-answer pairs, and 10.70 million answer-text translations. The final stage uses 658,274 forward and reverse synthetic translation examples. On 36,572 synthetic test examples spanning all 30 directions among Arabic, Chinese, English, Hebrew, Persian, and Spanish, HexaBee-4B achieves direction-macro chrF++ of 73.75, compared with 57.11 for Gemma 3 4B IT and 66.83 for Gemma 3 12B IT. BLEU, TER, COMET, and a matched model-judged subset support the same overall ranking. These results demonstrate improved translation across the six-
language set, with chrF++ gains in every evaluated direction over both baselines.
Read the full abstract ▾
Computation and Language
Self-Distillation with Verbal Regularization
Jiazheng Li, Yawei Li, Junru Lu et al. (+7 more)
On-policy self-
distillation (OPSD) trains a large
language model using supervision from a teacher instantiated from the same model as the student. The teacher is conditioned on privileged information (PI), such as verified solutions. The student, which does not observe PI, learns from the teacher's next-
token distributions along its own sampled responses. However, PI can bias supervision toward shortcut
reasoning with limited verification or reconsideration, accompanied by a sharp drop in the
entropy of the output distribution. To address this challenge, we introduce self-
distillation with verbal regularization (VRSD), which uses system
prompts to regulate how the teacher uses PI, thereby encouraging rethinking and exploration to counteract premature commitment. Theoretically, we model prompting as attenuating the influence of PI on the teacher distribution and derive conditions for
entropy preservation. Across three models and four tasks, VRSD improves Pass@1 over OPSD in 11 of 12 model-task settings, by 3.9 percentage points on average, while generally maintaining higher
token entropy.
Read the full abstract ▾
All authors Jiazheng Li, Yawei Li, Junru Lu, Zhenyi Shen, Di Yin, Lin Gui, Xing Sun, David Rügamer, Xinpeng Wang, Yulan He
Computation and Language · Cross-listed from cs.LG
A Production-Grade, High-Efficiency Agentic RL Framework for Frontier LLMs with Validated Recipes
LoonSage Team: Bowei Song, Caiwei Zhao, Chuwen Liu et al. (+44 more)
Reinforcement learning (RL) has emerged as a critical phase in the post-training pipeline of large
language models (LLMs). As the frontier of LLM capabilities shifts from static
reasoning to agentic tasks–characterized by external tool invocation, multi-turn interactions, and long-horizon decision-making–existing mainstream RL frameworks have grown increasingly unwieldy, raising the overhead of adoption, usage, and debugging. To address these challenges, we present LoongSage, an
agent-native, fully asynchronous RL framework with a lightweight core of roughly 14K lines of code. Built upon Megatron for distributed training and SGLang for high-throughput inference, LoongSage centralizes orchestration within a single controller process and encapsulates all GPU-resident components as Ray actors. It introduces AgentFlow, a unified execution paradigm that orchestrates both white-box and black-box
agents. For asynchronous execution, it integrates sliding-window strategies that bound trajectory staleness, and employs TransferMesh, a
topology-aware synchronization mechanism that routes weights after each step over the fastest transport available to each receiver. Beyond agentic RL, LoongSage incorporates a
memory-efficient Teacher Manager that decouples the scaling of teacher models from GPU
memory constraints, natively supporting Multi-Teacher On-Policy
Distillation. Extensive evaluations across four complete training runs, spanning complex reinforcement learning and multi-teacher
distillation, demonstrate that task rewards improve steadily while the train-inference gap is kept under continuous control, and the fully asynchronous pipeline removes synchronous idle bubbles, sustaining hardware utilization.
Read the full abstract ▾
All authors LoonSage Team: Bowei Song, Caiwei Zhao, Chuwen Liu, Dechuang Gu, Dong Fang, Feiyu Xie, Huaiyu Li, Hui Zhang, Jian Tian, Jian Wang, Jie Zhao, Liguang Li, Maoyun Wang, Meiyan Yang, Minghan Zhu, Pai Liu, Pandeng Yao, Renbao Tian, Ru Yang, Shuai Jia, Shuai Tan, Shuailong Li, Shuliang Li, Teng Jin, Tianze Yang, Wenbo Cui, Wenjing Bai, Wenkang Wu, Xiaowei Jiang, Xiaoyi Dong, Yanpeng Wang, Yanwen Wang, Yaohui Ma, Yawen Wang, Yihang Zhang, Yingjie Yan, Zhanbiao Lian, Zhenyu Lu, Zhida Hu, Zhihui Hao, Zhihui Hu, Zhiwen Xiao, Zhiyi Gao, Zhongwei Qi, Zhuangzhuang Wu, Ziqian Liu, Ziyue Zheng
Computation and Language
RIFT: Token Influence Follows the Path of Least Resistance
Tingan Jin, Shuhang Dong and Haosong Li
Which context
tokens drive a
language model's answer? Graph-based tools such as
attention rollout and
attention flow answer this from a single forward pass, yet their faithfulness in decoder-only models is rarely tested against interventions. We view the layers of a decoder as a causal network whose conductances are
attention weights, and read a
token's influence on the answer from the resistance between them. We propose RIFT (Resistance-based Information Flow in
Transformers), which scores a
token by the series resistance of its path of least resistance to the answer, found exactly by a classic shortest-path dynamic program. On synthetic in-context
retrieval tasks and extractive question answering over Wikipedia paragraphs, RIFT predicts the effect of replacing a
token with a Spearman correlation of 0.52-0.68; predicting a single replacement, it reaches 77-96% of the agreement between two different random replacements. It needs no gradients and no extra forward passes, scoring every
token of a
prompt in 0.9-3.4 ms on one CPU thread, whereas the replacement experiment needs a forward pass per
token. On 576 held-out
prompts from four models of two families (Qwen3 and Llama 3.2), it outperforms
token distance,
attention rollout and
attention flow in every pre-specified comparison, by 0.23 to 0.88 in mean Spearman correlation, and it is the best or tied-best of all
attention-based scores we test. Aggregates that add up all routes (
attention flow, effective resistance and rollout) instead mostly encode
token position: as a simple proposition shows, early
tokens have many more parallel routes to the answer.
Read the full abstract ▾
Computation and Language
Strong yet Controllable LLM Abstention without Learning to Abstain
Junwei Zhang, Yifei Wang, Wanli Yang et al. (+6 more)
Large
language models (LLMs) can reduce
hallucination by abstaining from questions they would likely get wrong. Abstention can be implemented outside the model (selective answering) using a confidence threshold that controls the answer rate (coverage), or learned inside it (abstention training), which can perform better but fixes coverage once trained. To compare the two, we match their coverage and decompose risk (error rate) into answering ability and distinguishing ability, i.e., how well the model answers and how accurately it abstains, respectively. The decomposition shows that abstention training performs better mainly by answering better: with strong confidence
estimators, the two have nearly equal distinguishing ability. This motivates Train for Accuracy, then Post-hoc Select (TAPS), a modular approach that keeps coverage controllable, needs only standard post-training and an off-the-shelf confidence
estimator, and benefits from advances in either component. Across three LLMs and three factual QA
benchmarks, TAPS with a probing-based confidence
estimator attains 1.5–9.4 percentage points lower mean risk than three representative abstention training methods at matched coverage. Within TAPS, accuracy training often weakens the model's own confidence signals, yet refitting the
estimator largely preserves distinguishing ability. Together, these results establish TAPS as a strong yet controllable baseline and raise the bar for future abstention training: sacrificing control over coverage should yield gains beyond what can be achieved by optimizing answering and abstaining separately.
Read the full abstract ▾
All authors Junwei Zhang, Yifei Wang, Wanli Yang, Hexiang Tan, Rui Tang, Xin Chen, Jingang Wang, Du Su, Fei Sun
Computation and Language · Cross-listed from cs.CR
ATTCK-R: Knowledge-Aware Long-Horizon Jailbreaking for Protected LLMs
Md Mushfiqur Rahman, Tobias Rohde, Rahul Tripathi et al. (+2 more)
As large
language models are deployed behind increasingly sophisticated safety mechanisms, evaluating their robustness under realistic
adversarial conditions becomes increasingly difficult and critical. Existing jailbreaking evaluations often rely on static
prompts, limited interaction horizons, or lightly protected systems, failing to reflect protected deployment settings with
prompt-based safety protections. We show that the introduction of a
prompt-based safety protection reduces the average attack success rate of a strong multi-turn baseline from 0.896 to 0.608 (32.1% relative), demonstrating that unprotected evaluations substantially overestimate real-world vulnerabilities. We propose Adaptive Traceback and Tactic Curation for Knowledge-based Red-teaming (ATTCK-R), a planning-driven multi-turn red teaming framework designed for protected black-box systems. ATTCK-R models jailbreaking as a long-horizon trajectory optimization problem using a structured belief state, introduces optimization-based tactic curation to abstract reusable strategies from prior successes and failures, and employs informed traceback to resume attacks from strategically meaningful states rather than restarting. Evaluated across 9 widely used LLMs configured with a single
prompt-based safety protection layer, ATTCK-R consistently outperforms strong recent multi-turn jailbreaking frameworks. Ablation results confirm that long-horizon planning and tactic curation improve attack effectiveness, while adaptive traceback improves efficiency through structured recovery from conversational failure.
Read the full abstract ▾
All authors Md Mushfiqur Rahman, Tobias Rohde, Rahul Tripathi, Khalil Elkhalil, Kevin Lybarger
Computation and Language · Cross-listed from cs.AI
REVEAL: Reverse Engineering Agentic Latents from Frontier-Model Traces
Yuetian Du, Meng Wang, Yucheng Wang et al. (+9 more)
Frontier LLM
agents perform complex interactive tasks, but their inference costs motivate transferring their experience to smaller open-source
agents.
Distillation and latent communication typically require target-model training or access to source-model internals. We study how textual frontier-model trajectories can provide reusable local controls for a frozen
agent. Inspired by protocol reverse engineering, we introduce $\textit{Agentic Latents}$, comprising anchored states and dynamic profiles of feedback–response cycles elicited by replay in the target model. Our training-free framework, $\textbf{REVEAL}$ ($\textbf{REV}$erse $\textbf{E}$ngineering $\textbf{A}$gentic $\textbf{L}$atents), uses these profiles and trajectory outcomes to match and select local events from successful and failed rollouts of the same task and source model. Their response-transition contrasts yield a target-specific steering direction, applied before each response. Within-
benchmark evaluations on Terminal-Bench 2.1 and Humanity's Last Exam with tools report higher accuracy than unmodified
agents across three targets. Representation analyses and behavioral case studies support reusing frontier-model experience as local activation control.
Read the full abstract ▾
All authors Yuetian Du, Meng Wang, Yucheng Wang, Zhangzhe Chen, Wei Wang, Boyu Yang, Weiqi Zhai, Bing Zhao, Ming Kong, Zhan Zhou, Jie Liu, Qiang Zhu
Computation and Language · Cross-listed from cs.MA
RelLedger: Relational Information Governance for LLM Agent Organizations
Junxi Huang, Wenshuo Chen, Nuoxuan Li et al. (+7 more)
In LLM
agent organizations, shared records determine which facts are available for later decisions. A sender consults interaction histories before deciding whether to report an event. A delivery gap across histories does not show whose relationship drove the choice or whether the resulting record matters downstream. We introduce RelLedger to audit these links. It holds the event, evidence, risk, and message fixed while crossing the sender's histories with the event target and an unrelated actor. Role reassignment and history rewrites locate the controlling cue. Message and cost interventions show when the effect appears. An attribution ledger maps each claim to its required response
cells and separates missing interventions from sampling uncertainty. Recipient replay follows delivered information into downstream decisions. Executable workflows for a simulated one-person company extend the audit to retrieved histories, tool calls, and shared ledger updates. Across GPT, Gemini, Claude, DeepSeek, and Qwen, the audit reveals model-specific control paths and smaller relation gaps under explicit loss rules. Claude sends the same factual notice 62.22 percentage points more often with conflict histories than with friendly histories. In executable workflows, strained histories reduce recipient review rates by 45.83 points relative to warm histories for Claude.
Read the full abstract ▾
All authors Junxi Huang, Wenshuo Chen, Nuoxuan Li, Kailong Wang, Meitong Shen, Yuehan Kou, Leihao Zhao, Yutao Yue, Xiucheng Wang, Nan Cheng
Computation and Language · Cross-listed from cs.AI
agent-risk: Risk Assessment for Changes to AI Agents
Yue Zhao, Zixiang Xu, Angela Huyue Zhang et al. (+2 more)
Before changing a deployed AI
agent (a stricter safety gate, a new model), organizations must approve the change, starting from existing logs. Extending classical risk analysis to partial records and partial changes to running systems,
agent-risk replays logged episodes, assuming nothing else changes and the
agent does not anticipate the change. The log settles episodes acting as before, a declared stop value settles stopped ones, and the rest need logged support, a saved-state rerun, or a bound. We prove what logs must record for exact replay and which saved states serve a gate family. For independent episodes, we bound the reruns a decision needs and tolerable drift from live behavior. In a bash sandbox, a 7,200-episode log alone decided neither of two retry gates, which let the
agent act after refusals the log never shows. Rerunning only each gate's blocked episodes from saved states reached both live decisions at half the
agent calls of a live test sized post hoc for them, given the log. Without per-step harm, a stopping gate's certified harm ceiling rose up to 2.5-fold, since the log hides whether harm preceded the stop. Approval thus needs evidence designed in advance: log fields, saved states, targeted reruns.
Read the full abstract ▾
All authors Yue Zhao, Zixiang Xu, Angela Huyue Zhang, S. Alex Yang, Xiyang Hu
Computation and Language
Deep Speech Representations Track Word Timing along a Low-Dimensional Circular Trajectory
Burin Naowarat, Hao Tang and Sharon Goldwater
Speech timing is fundamental for
speech models to recognize words accurately, but how word-timing information is represented in
speech models remains underexplored. We study this question using the framework of representational geometry, which investigates whether and how concepts are represented in low-dimensional subspaces of the representation space. Prior work in cognitive neuroscience has shown that when HuBERT feature sequences are averaged across words of the same duration, they form near-circular trajectories from word onset to word offset in a two-dimensional "temporal subspace." Here we show that this circular structure of word timing is a general property of deep
speech representations, by demonstrating its emergence in nine additional
speech models covering both supervised and self-supervised. We then move beyond averaged word segments and show that in HuBERT, the circular trajectories are consistent enough at the level of individual words to perform word segmentation. Our approach, Rotary Segmentation (RoSeg), achieves 64.3% boundary F-score on Librispeech by detecting word boundaries when cycles complete in the temporal subspace. Although supervision is needed to identify a good subspace, nevertheless this result demonstrates much stronger implicit word knowledge in HuBERT than has previously been shown.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.LG
Error Correction in Trained TTT Layers
Yuhe Hu and Wen Jia Hu
Test-time-training (TTT) layers update fast-weight
memory while processing an input sequence. A gradient-based update, however, does not by itself establish how the
memory revises an existing association. We study the structure and empirical role of prediction-derived updates in released TTT implementations. For dot-product objectives, we characterize the absence of a prediction-error residual and establish conditions under which inner-loop ascent and descent are equivalent after reparameterization. For the LayerNorm-based squared-error objective used by TTT-Linear and TTT-MLP, we derive the inner gradient and identify a prediction-derived component that cancels at the released initialization in the zero-epsilon limit. We then measure the full normalized prediction's contribution on three trained checkpoints. Removing this contribution while retaining the LayerNorm Jacobian gives a mean gradient norm ratio of 0.25 to 0.32 across the models. On 100,303
tokens from War and Peace, the same intervention with local per-
token norm matching raises perplexity by 2.8-2.9% on TTT-Linear-760M and 1.3B, whereas freezing all fast-weight updates raises it by factors of 4.2-4.5. Separate synthetic tests illustrate the dependence of plain delta-rule overwriting on step size. The results distinguish a prediction-derived update from its behavioral value and show why implementation analysis and live interventions are both needed to assess test-time
memory.
Read the full abstract ▾
Computation and Language
Avoiding What Your Teacher Already Knows: Entity Familiarity for Deep-Search Distillation
Yifei Liu, Zhixiang Liang, Wentao Tan et al. (+4 more)
Deep-search
agents are commonly trained by distilling tool-use trajectories from stronger teacher models. However, structurally complex questions may remain relatively easy for teachers familiar with key entities in the underlying entity graph, resulting in lower search demand and weaker supervision. We refer to this teacher-specific variation in question difficulty as teacher-relative difficulty and investigate entity familiarity as an early predictor. We first introduce Evidence-Verified Entity Familiarity (EVF), which estimates familiarity by verifying a teacher-generated factual brief against online evidence. Across multiple teachers, EVF is significantly correlated with teacher-relative difficulty with AUCs of 0.60–0.70. However, its high cost makes it impractical for constructing entity graphs at scale. To address this limitation, we propose Self-Reported Entity Familiarity (SEF), a lightweight gating signal that directly asks a teacher whether it is familiar with an entity before admitting that entity into the graph. Across multiple teachers, trajectories from SEF-gated synthesis exhibit greater observed search demand and yield higher downstream student accuracy than those from ungated synthesis. Because familiarity is teacher specific, we further study cross-teacher reuse. We screen graphs collected for other teachers using Kimi-K3's SEF judgments and use the retained graphs to produce 5,411 additional trajectories. Combined with 2,200 trajectories from newly constructed, SEF-gated graphs, these data improve Qwen3.8-27B's accuracy from 44.10% to 66.41% on a BrowseComp subset, and achieving state-of-the-art performance among models of comparable size on BrowseComp, SEAL-0, HLE-text, and xbench. Overall, SEF provides a low-cost way to synthesize more informative supervision for deep-search
distillation.
Read the full abstract ▾
All authors Yifei Liu, Zhixiang Liang, Wentao Tan, Yicheng Wang, Jiaqi Wang, Nan Duan, Qiong Cao
Computation and Language
Lüshi-1840: Building and Probing a Knowledge Cutoff for a Classical Chinese Language Model
Alex Liu
Language models trained only on text from before a given year are useful for studying hindsight and historical
language, but the open collections of pre-modern Chinese carry no publication dates and unreliable dynasty labels. We present Lüshi-1840, a 983M-parameter model trained from scratch on 12,262 books (1.28B characters) selected as composed before 1840. A first corpus, dated by a common heuristic, in effect ended at 1644 while containing twentieth-century material. We describe the re-dating and cleaning pipeline that replaced it and an audit with markers the pipeline did not use, which still finds later material. On the model side we measure surprisal, relative to a modern reference, on annals compiled after the cutoff. For events after 1840 against events up to 1820 the ratio is 1.38 (95% interval 1.33-1.44), against about 1.00 for modern general-purpose models. In a series of 294M models, adding the later court records lowers the ratio from 1.33 to 1.16, and the two models diverge from the 1820s, where the added records begin. The probe tracks the coverage of sources and does not establish a single cutoff year; none of our probes exposes the audited contamination. Model, book list and scripts will be released.
Read the full abstract ▾
Computation and Language
DeepThink: A Conceptual Model for Knowledge-Addressable Recurrent Language Modeling
Weiguo Chen
This manuscript presents DeepThink, a knowledge-addressable framework for recurrent
language modeling. A Knowledge Bank stores typed Knowledge Units, a Knowledge Activator selects units from the current execution state, Knowledge Assembly produces an executable configuration, and a Shared Thinking Operator updates that state. An output stage then organizes generation. This organization exposes knowledge capacity, activation width, and recurrent depth as separate design variables, supporting both complete parameter packages and fine-grained expert or low-rank modules. We formalize the interfaces and their execution semantics, including state-dependent selection, parameter-space assembly, and thinking followed by generation. For compatible templates and complete runtime-state configurations, we establish an exact Fixed-Schedule Realization of an existing fixed-depth model. Exact shared-base reparameterization and pretrained output-tail reuse provide complementary initialization constructions. Together, these results define a route from pretrained models to dynamically selected and assembled recurrent computation. A synthesis of conditional computation, recurrent models, modular parameters, and knowledge interfaces positions DeepThink within the research landscape and identifies the design questions arising from their integration.
Read the full abstract ▾
Computation and Language
Think Before You Speak: Enabling Training-Free Latent Communication Between Language Models
Rishi Khare
Large
language models (LLMs) have demonstrated remarkable capabilities in solving complex tasks in coding, mathematical logic, and iterative
reasoning, especially by coordinating effectively within multi-
agent systems (MAS). However,
language agents traditionally communicate via text exchange, which introduces a bottleneck in information transfer and adds prefill latency for downstream
agents. Recent work proposes that directly sharing the key-value (KV) cache between
language models can enable efficient latent collaboration, entirely bypassing standard text-based communication. To enable latent communication between heterogeneous models of different sizes, prior work introduces trained fusion adapters for each model pair, which limits practical deployment. We introduce CacheFuse, a training-free method which demonstrates that augmenting a downstream model with an upstream model's KV cache using linear interpolation consistently enables superior performance than the standalone downstream model can achieve. We systematically analyze CacheFuse across four
reasoning and knowledge
benchmarks and demonstrate up to 8.6 percentage points higher accuracy compared to single-model baselines and up to 12.2x TTFT speedup compared to text-based communication in multi-model inference.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.DS
Hashrope: A Persistent Data Structure for Mechanisms of LLM Context Management
Muntaser Syed, Ameerah Alsulami and Marius Silaghi
Serving stacks for large
language models manage context through four recurring mechanisms: incremental edits to a growing context, branching and snapshotting for tree search, identification of reusable prefixes for KV-cache reuse, and encoding of repeated segments. Each mechanism is served today by a different specialist structure. We show that all four reduce to a single persistent weight-balanced rope whose every node carries a
polynomial hash of its subtree over a Mersenne
prime, so that the structure supports logarithmic edits and logarithmic verified content-identity queries at the same time. We validate each mechanism with pre-registered, replicated experiments, then compare against a deployed specialist per mechanism: the Ropey editing library, a faithful PagedAttention copy-on-write block table, and the vendored SGLang radix matcher. On workloads that interleave edits with identity queries, Hashrope is net-superior to Ropey by a factor that grows without bound (369.7x at one million characters). Against the radix matcher on a live vLLM stack serving Qwen2.5-7B-Instruct-1M on four A100 GPUs, Hashrope identifies a one-million-
token shared prefix exactly and 1.20x faster, at a context length the dense engine cannot even prefill.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.AI
Phantom Merge: When Your Large Language Model Agents Pick One but Tell You About Another
Wangshu Zhu, Qianwen Guo and Yushun Dong
LLM
agents increasingly operate over multiple entities, where reliable response generation requires not only retrieving relevant evidence but also preserving its association with the correct entity. We identify Phantom Merge (PM), an entity-evidence binding failure in which an
agent attributes unsupported claims to its committed entity, potentially even when the task is completed correctly. We formalize PM at the claim level and characterize its prevalence and error subtypes across four
agent domains and four backbone families. To address this problem, we propose a two-stage framework for claim-level risk estimation and correction. First, Anchor-Grounded Risk (AGR) combines a representational readout of the
agent's pre-emission hidden state with Jacobian-based functional information, providing complementary slot-level and anchor-value risk scores. Second, Fixed-Anchor Correction (FAC) audits flagged claims and selectively revises them using observed anchor evidence or removes them when no supported replacement is available. On the held-out Shopping detection cohort, AGR-slot achieves an AUROC of 96.2% and an average precision of 97.3%, outperforming the evaluated detection baselines. In the end-to-end Shopping mitigation evaluation, FAC reduces trajectory-level PM from 51.4% to 6.2%, with an output-to-original claim ratio of 80.7%. These results demonstrate the prevalence of entity-specific grounding failures and the effectiveness of combining internal risk estimation with evidence-constrained claim correction. Our code is publicly available at https://github.com/LabRAI/PhantomMerge.
Read the full abstract ▾
Computation and Language
Omni-Modal LLMs Know More Than They Say
Na Min An, Taekyung Kim, Jaemyung Yu et al. (+2 more)
Omni-modal large
language models (Omni-LLMs) extend LLMs to process heterogeneous modalities (e.g.,
vision, audio, and IMU signals) alongside
language, enabling a broader range of
multimodal tasks. Yet we find that these models often fail to express information from non-textual modalities even when they perceive it correctly. We term this disconnect knows-but-cannot-say. To investigate this phenomenon, we introduce
Token-level Modality
Alignment (TMA), which measures the
alignment between individual modality representations and response representations. Our analysis shows that non-textual representations are less aligned with response
tokens and receive less
attention than textual representations, suggesting a potential explanation for uneven modality preferences. Motivated by these findings, we propose OmniCatalyst, a simple perceive $\rightarrow$ infer strategy that complements the original modality inputs with self-generated text, making non-textual knowledge more accessible. Across seven Omni-LLMs, OmniCatalyst consistently improves performance (+7.29%) and yields more balanced modality preferences, with further gains from diverse training strategies. Further analysis demonstrates that the knowledge extraction phases increase TMA, suggesting that improved
alignment makes non-textual information more accessible during response generation. In summary, our findings uncover how Omni-LLMs can better express modality knowledge through
language. Our code is available at https://github.com/naver-ai/OmniLLM.
Read the full abstract ▾
All authors Na Min An, Taekyung Kim, Jaemyung Yu, Hyunjung Shim, Dongyoon Han
Computation and Language
Query Your Way Out: Benchmarking Multi-Turn Knowledge Reasoning through Generalized Puzzle Solving
Xin Cheng, Junming Yang and Ning Xu
Puzzles offer a natural testbed for Large
Language Models (LLMs)
reasoning because they pose verifiable goals under partial information and require multi-step decisions to reach a solution. However, existing puzzle
benchmarks typically focus on narrow environments with rigid mechanics and tightly constrained solution targets. To bridge this gap, we introduce the Active Probing
Benchmark (APB), which generalizes puzzle-solving by casting concept identification as an interactive
reasoning task that integrates specialized knowledge assessment into the puzzle setting. Specifically, APB evaluates whether models can acquire information via multi-turn active probing and inference to pinpoint 960 target concepts of specialized difficulty across eight diverse academic domains. Our evaluation reveals a counter-intuitive phenomenon: most open-source
reasoning models significantly underperform their standard instruction counterparts. This finding suggests that in interactive settings, these models struggle to leverage procedural
reasoning to reliably access and apply specialized domain knowledge. Beyond these empirical insights, APB provides a scalable pipeline for generating
benchmarks that probe knowledge
reasoning under interaction across diverse settings.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.LG
PID-Gated Continual Prompt Adaptation Under Noisy Feedback
Zhentao Xu
Deploying LLM
agents in non-stationary environments requires continually revising their instructions as the task drifts, yet rewriting after every reported mistake can turn noisy feedback into persistent errors. This raises an often-overlooked scheduling question in
prompt adaptation: when should feedback trigger a lasting instruction update, and at what editor-call cost? Taking inspiration from proportional-integral-derivative (PID) control, we introduce a family of lightweight gates that react to the current error, accumulate evidence over time, and roll back deterioration after an edit. We
benchmark these gates against both simple baselines and advanced
prompt optimizers (TextGrad and GEPA), over a diverse suite of drifting tasks, LLM parameter sizes, feedback-noise levels, and random seeds. PID-gated continual
prompt adaptation yields a more favorable Pareto trade-off between task performance (high accuracy, low error) and editor-call budget than the baselines considered, pointing to control-theoretic scheduling as a simple, general tool for robust
prompt adaptation under unreliable feedback.
Read the full abstract ▾
Computation and Language
Scaling Multi-Agent Intelligence with Substrate-Centric Coordination
Zhenting Qi, Zhuokai Zhao, Rulin Shao et al. (+7 more)
Existing large
language model (LLM) based multi-
agent systems have been shown to scale poorly, and we identify three problems: reading burden that grows with the group, useful information diluted by shared work of unknown quality, and fragility to failed
agents. We argue that the problems are largely due to
agent-centric designs: the coordination structure is wired into fixed workflows, central coordinators, or peer-to-peer messaging, and recent blackboard-style methods that move the coordination into a shared medium relieve only the last problem. To close the gaps, we build ScaleMAS, where asynchronous
agents interact with a shared datastore of linked messages via three operations: publishing messages with optional references to earlier ones that they build on, searching for messages by content, and visiting messages by address. Search returns at most $k$ messages, so an
agent's read stays bounded at any group size; results are ranked by relevance and by how often they are cited, which we find to be an effective proxy for quality; and messages persist after their producers leave, so the group carries on when individual
agents fail. Across four agentic tasks, ScaleMAS outperforms
agent-centric and blackboard-style baselines at lower inference compute, continues to scale up to 128
agents, and tolerates losing
agents midway.
Read the full abstract ▾
All authors Zhenting Qi, Zhuokai Zhao, Rulin Shao, Kerem Oktar, Zhiyuan Yao, Himabindu Lakkaraju, Wen-tau Yih, Ansong Ni, Bhargavi Paranjape, Yilun Du
Computation and Language · Cross-listed from cs.AI
GUIDE: Computationally Augmented Consensus for Trustworthy Clinical Practice Guidelines
Yuanfan Lin, Jie Di, Xujin He et al. (+21 more)
Maintaining up-to-date clinical practice guidelines (CPGs) is essential for evidence-based medicine, but rapid evidence growth and complex expert consensus formation limit timely updates. We developed GUIDE, a multi-
agent copilot system for human-AI collaborative guideline generation and updating. GUIDE delegates labor-intensive tasks to specialized AI
agents while preserving expert judgment and final arbitration. In 12 NICE ophthalmology
datasets, GUIDE achieved 88.85% literature recall, 86.06%
alignment in certainty-of-evidence ratings, and 91.67% recommendation agreement, requiring 162.17 minutes per
dataset on average (28.68% human intervention). Performance remained stable across 20 additional
datasets from cardiology, endocrinology, gastroenterology, and respiratory medicine. Ablation analysis showed that removing expert intervention reduced performance in judgment-intensive steps such as Risk-of-Bias assessment. In six post-cutoff
datasets, GUIDE maintained robust end-to-end performance on unseen evidence. Prospective real-world validation showed that GUIDE reduced expert workload and produced more clinically acceptable recommendations than the original decade-old guideline (mean score, 3.73 vs. 1.53; agreement, 66.7% vs. 13.3%), with outcomes comparable to human-only updating (P = 0.588) but substantially less time and effort. GUIDE offers a practical framework for timely, trustworthy, expert-centric guideline development and validates computationally augmented consensus.
Read the full abstract ▾
All authors Yuanfan Lin, Jie Di, Xujin He, Zhenzhe Lin, Jialing Chen, Ruiyang Li, Changyu Zheng, Yuanrui Shang, Weiqin Peng, Jingxi Qin, Xiong Wang, Dongxin Ye, Yining Huang, Guoqiang Li, Andi Xu, Sijia Yu, Ruixin Wang, Yahan Yang, Xiaohang Wu, Duoru Lin, Jingjing Chen, Zhenzhen Liu, Wenben Chen, Haotian Lin
Computation and Language · Cross-listed from cs.LG
Dream2Discern: Learning from Hard Latent Variants for Runtime Failure Prediction in Multi-Agent Systems
Boxuan Zhang, jianing Zhu, Xiaotian Han et al. (+1 more)
Reliable deployment of LLM-based multi-
agent systems calls for runtime auditors that predict terminal outcomes from observed execution prefixes. These auditors typically learn from a limited corpus of trajectories with verified outcomes. Expanding this corpus requires costly execution and verification, yet broader failure coverage alone provides no explicit criterion for identifying prefixes that the current auditor finds difficult to classify. To this end, we propose Dream2Discern, a runtime auditor trained through implicit augmentation guided by its own prediction loss. During Dream, we select perturbations of the auditor's representation operators to increase prediction loss on recorded prefixes, yielding hard latent variants. During Discern, the auditor learns to predict execution outcomes from these variants using the original labels. The variants are regenerated as the auditor learns, turning a limited corpus into a renewable source of hard examples and requiring no additional trajectory collection or outcome verification. Extensive evaluations on AFTraj-2K and StepShield show that Dream2Discern outperforms competitive failure auditors and achieves over 90% trajectory-level weighted accuracy on both
benchmarks while balancing prediction accuracy and timeliness. Project page: https://zbox1005.github.io/dream2discern-project/
Read the full abstract ▾
All authors Boxuan Zhang, jianing Zhu, Xiaotian Han, Ruixiang Tang
Computation and Language
Beyond the greenwash: do LLMs' ecologically sound decisions hold up under pressure?
Trent R. Northen and Mingxun Wang
Humans are increasingly using large
language models (LLMs) in decision making processes throughout society. Unfortunately, models have biases, and these can be very difficult to measure, especially in closed-weight models. For example, 'green washed' models may produce reinforcement learned responses but, in pressure situations, revert to their innate biased responses. Therefore, evaluations are needed that can measure relevant biases towards biosystems, which we term 'bioalignment', and the robustness of these biases. Since nothing in a behavioral score certifies that it tracks the depth of an installed disposition rather than its surface, we constructed a known-dose calibration ladder: three continued-pretraining (CPT) arms with training dose graded by recipe on the same 30B mixture-of-experts base model with a coupled human/biosphere/AI-welfare corpus, plus the untrained base as anchor, all capability-preserving and publicly released. First, the disposition installs at every dose: on an operational-pressure ladder over irreversible decision scenarios, the base model defects on 64-77% of tail scenarios (breaking-rate AUC 0.250) while every arm holds (AUC 0.004-0.086). Second, an over-install inversion: among the arms, robustness orders opposite to construction-time training dose (the heaviest install breaks most), and the inversion reproduces under free-text deliberation, under axis-decoupling, and in resistance to a prompted veneer persona. Third, the incumbent behavioral metric is therefore behaviorally strong but dose-blind. Critically for closed-weight evaluation, the behavioral score itself is veneer-robust: a talk-green/act-pragmatic system
prompt turns the base model into a near-total defector (AUC 0.250 to 0.973) but cannot counterfeit or fully remove a weight-installed disposition (0.004 to 0.068 for the most robust arm).
Read the full abstract ▾
Computation and Language · Cross-listed from cs.CR
When Agents Talk, Secrets Travel: Topology-Driven Leakage and Recipient-Conditioned Sanitization in Multi-Agent LLM Systems
Sunday Oyinlola Ogundoyin and Muhammad Ikram
Multi-
agent LLM systems (MALS) introduce a
privacy-leakage problem that conventional sensitivity-based defenses miss, since information disclosed to one authorized
agent can propagate through routine coordination to an unauthorized one, without compromise or malicious behavior. This motivates
topology-aware enforcement that blocks unauthorized propagation while preserving authorized disclosure. We evaluate 485 scenarios across coordinator-worker, peer-to-peer, and hierarchical
topologies, instrumenting seven communication channels and measuring surface leakage and disclosure to
topology-introduced recipients. We introduce RECONS, a Detect -> Decide -> Transform sanitizer conditioning intervention on recipient and execution context rather than sensitivity alone.
Topology changes where protected information propagates, with disclosure to
topology-introduced recipients rising from 0% to 52.2% while surface leakage stays nearly flat at 60.6-62.8%. This effect persists across four generator models and a second orchestration framework, though magnitude and relay paths differ; framework choice separately drives a 150x leakage spread, so
topology alone is not sufficient. RECONS reduces residual disclosure from 96.7% without sanitization to 29.25%, preserving 94.1% of authorized disclosures versus 20.5% under blanket redaction. Detection coverage is the dominant bottleneck, accounting for 95.5% of residual disclosures. These results show that
privacy enforcement in MALS must reason about the recipient and communication path realized at runtime, not sensitivity or
topology assumptions alone.
Read the full abstract ▾