Computation and Language
Shared Stopping Decisions Change Answers in HQQ Cache Quantization
Seunghui Jwa, Minsu Oh, Chanjun Park et al. (+1 more)
Language-model systems batch questions for throughput, but unrelated questions should not change a target's answer when its input and numerical execution are fixed. We study compression of the key and value cache, which stores
attention representations reused during generation. With request-local groups,
Transformers' Half-Quadratic Quantization (HQQ) backend updates compression parameters separately but uses a shared average error to decide when all updates stop. Replacing only the question batched with the target changes four-bit HQQ answers in 170/384 test comparisons across two models. Replaying the other execution's update counts reproduces its complete answer and cache fingerprints in every changed pair, in both directions. Computing the stopping mean in FP32 reduces cache differences but leaves answer changes. Native HQQ also changes confirmed numerical correctness in eight arithmetic pairs. Fixed iterations and request-local stopping remove observed companion dependence under matched controls. Request-local stopping remains sensitive to synthetic padding changes at the tensor level. Fixing the original iteration budget removes this decision path without tuning. Neither repair has an established quality advantage, and natural rebatching still changes answers. Request-independence audits must cover stopping decisions as well as quantization groups.
Read the full abstract ▾
All authors Seunghui Jwa, Minsu Oh, Chanjun Park, Yeo-Chan Yoon
Computation and Language · Cross-listed from cs.CR
DP-ES: Differentially Private Evolution Strategies for Prompt Optimization
Ziniu Liu, Aiping Li, Yue Han et al. (+5 more)
Token-level differentially private (DP)
prompt optimization methods such as DP-OPT can become unstable under tight
privacy budgets: on GSM8K, DP-OPT obtains $49.5\pm28.5\%$ across 30 runs, and a logged search trajectory reveals
prompt-template drift and noise-sensitive irreversible choices. We diagnose these as structural consequences of greedy
token-by-
token construction over privately aggregated counts. We then propose DP-ES (Differentially Private
Evolution Strategies), a structurally cleaner alternative that maintains a population of full
prompts, mutates them via LLM calls that never access the private
dataset, and spends
privacy only on sampled-Gaussian evaluation; deterministic or Gumbel-smoothed selection is post-processing. Under a conservative $(\varepsilon\leq1.0,\delta=10^{-5})$ guarantee, DP-ES achieves 88.1% on GSM8K (+38.6 pp over DP-OPT, approximately 9 times lower standard deviation), 99.7% on MedQA, 73.5% on BANKING77, and 86.8% on Alpaca. It is also 2.5 times faster in wall-clock time and uses 3.3 times fewer logged private-data call groups than DP-OPT. Selection and population ablations, implementation-level noise checks, and a 200-profile exact-match memorization stress test complement the formal guarantee. Scope: Our experiments establish optimization robustness under DP noise, especially where
prompt structure is critical; end-to-end validation on genuinely sensitive, non-saturated deployment data remains future work.
Read the full abstract ▾
All authors Ziniu Liu, Aiping Li, Yue Han, Han Yu, Junjian Zhang, Dong Zhu, Changjian Li, Shiqiang Zhang
Computation and Language
Cross-lingual Calibration of Pre-Generation Success Probes for Multilingual LLM Routing
Andrea Paganelli, Stefano Civelli, Pietro Bernardelle et al. (+1 more)
Pre-generation success probes estimate response correctness from a
language model's hidden activations before decoding, enabling cost-aware routing. While prior work has demonstrated their utility primarily on English inputs, we study their reliability across
languages along three dimensions: (1) whether they preserve the ranking of likely successes and failures (DISCRIMINATION); (2) whether they retain probabilities that match observed success frequencies (CALIBRATION); and (3) whether they produce scores comparable enough across candidate models for cost-aware multilingual routing (UTILITY). Using 3,000 MATH problems in 10
languages and 8 open-weight model configurations, we compare cross-lingual transfer from English-trained probes and equal-budget pooled multilingual probes. English-trained probes retain useful cross-lingual discrimination but become less well calibrated after transfer. Pooled multilingual supervision improves both properties and yields more reliable estimates of success. In routing experiments, the pooled router achieves a 0.7% higher test success rate while reducing modeled cost by 13.0% relative to always selecting the model with the highest average success. These results show that multilingual routing requires success estimates that remain well calibrated and comparable across
languages and models.
Read the full abstract ▾
All authors Andrea Paganelli, Stefano Civelli, Pietro Bernardelle, Gianluca Demartini
Computation and Language · Cross-listed from cs.AI
Do Small Language Models Learn to Negotiate? A Controlled Scaling Study of RL-Trained Sellers
Pedro Tabacof and Sagar Joglekar
LLM
agents are starting to own the full customer experience. Soon, LLMs may be selling and buying on behalf of companies and customers respectively. Small models are more cost-efficient at scale, but can reinforcement learning train them into competent sellers? We train four Gemma 4 checkpoints (2.3B to 31B effective parameters) with GRPO on a programmatic utility reward for bilateral multi-issue bargaining, and evaluate every arm on the same 1,152 negotiations against two frontier buyers it never saw in training. With the same learning rate ($10^{-6}$) for every size, the gain of the RL model over its base rises from $+0.001$ at 2.3B to $+0.078$ at 31B. Each size was trained once and the two smallest checkpoints use a different architecture, so we fit no scaling law. Tripling the learning rate, with the same or fewer training steps, improves on the shared rate at every size by $+0.032$ (2.3B) to $+0.081$ (4.5B). In exploratory comparisons with two frontier models run as sellers, the 12B seller trained at the tripled rate scores above both, though its untrained base already scores as high as they do. The 4.5B seller at that rate shows no detectable difference from either and fits on one 48 GB GPU. A further 2.3B arm at ten times the shared rate raises pooled score, but its gain concentrates on the evaluation buyer that shares a model family with the training pool. These results suggest tuning the learning rate before concluding that a small model cannot learn to negotiate, and testing against buyers from more than one model family.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.AI
Judged Useless, Queried Anyway: Tool-Using Agents Rarely Turn Their Own Evidence Judgments into Stopping Decisions
Chubin Zhang, Zhenglin Wan, Xingrui Yu et al. (+4 more)
An
agent whose tool keeps returning nothing useful should stop relying on it. In a
retrieval environment with controlled source failures, we separate how
agents judge results from what they do. We compare stopping at the same step after longer and shorter runs of results the
agent judged useless; this contrast is zero for clock- or deadline-driven stopping. Where we record their judgments, the seven
agents we test call a failing source's results useless 97-100% of the time, yet most of them rarely stop on that judgment.
Prompt cues change when they stop but not what they stop on. Permission to answer from
memory and a
reasoning mode can bring early stops regardless of evidence, a stated budget moves the 7-8B models' stops to the deadline, and a stopping rule or call cost in the
prompt is followed at most partly. Stopping follows the evidence only when the harness enforces an integration step that makes the
agent answer after five consecutive results it judged useless. This step raises failing-source success for every model, keeps the stopping point fixed when the budget doubles, and needs no extra judgment call when the
agent states its judgments. A pre-registered replication on 300 fresh questions confirms the dissociation and the rule's effect.
Read the full abstract ▾
All authors Chubin Zhang, Zhenglin Wan, Xingrui Yu, Jingxuan Wu, Yaxin Zhou, Ivor Tsang, Bo An
Computation and Language · Cross-listed from quant-ph
What Does It Cost to Simulate a Quantum Sentence Classifier? An Energy and Compute Perspective on Near-Term QNLP
Kishlay Kashyap and Sandipan Ganguly
Near-term
quantum natural
language processing (QNLP) experiments often run on classical simulators, so simulator cost is part of the field's practical compute burden, yet accuracy tables do not show it. We measure that cost for a variational
quantum classifier (VQC) on binary SST-2 sentiment classification, using PennyLane's state-vector simulator over a controlled grid of 27 configurations: three balanced training-set sizes (N = 200, 500, 1000), three
qubit counts (4, 6, 8), and three
circuit depths. Each VQC is compared with logistic regression on the same PCA-reduced input; full TF-IDF logistic regression gives an uncompressed reference. The VQC beats its matched baseline in 6 of 27 single-seed comparisons. After reruns at two further seeds, only 1 of these 6 keeps a positive mean advantage larger than its paired seed-to-seed variability, and paired tests on the fixed validation set do not establish it. VQC training is 886-21,127 times slower in measured wall-clock time than the matched classical fit (median 3,158 times); going from 4 to 8
qubits roughly doubles simulator time, and within the tested range per-step cost is well approximated by a linear function of the parameter count. CodeCarbon energy and CO2 estimates are secondary: they imply an almost constant power of about 41 W, so they add little beyond runtime, and we do not build an energy ratio from them. The study is narrow (one
dataset, representation, ansatz, simulator, and CPU environment) and is a reproducible feasibility measurement, not a general verdict on QNLP.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.LG
Anosognosia in LLMs: Probing Self-Awareness of Quantized Computational Substrate
Yoshihiro Izawa, Gouki Minegishi and Yoko Yamakata
Can LLMs recognize degradation in their own computational substrate? Inspired by anosognosia, a neurological condition in which patients fail to recognize impairments in their own abilities, we investigate whether LLMs can recognize degradation in their computational substrate induced by quantization. We first show that existing models fail to self-report their quantization state, even when provided with their own generated text as an external cue. Linear probing reveals that, while generated text carries almost no trace of quantization, internal representations contain clear, method-specific fingerprints. Through training, models learn to identify severely degraded outputs such as those of 4-bit models by comparison, yet still fail to do so from a single output. A shared LoRA trained jointly across quantization levels succeeded in reading out internal fingerprints, but fails on unseen quantization methods, merely mapping method-specific fingerprints to labels. Whereas external self-observation can restore awareness in some cases of human anosognosia, our results suggest that the more promising route to enabling such awareness in LLMs may lie in their internal representations. Our results highlight fundamental limits of generalizability to LLM self-monitoring.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.AI
MS-Exam-Gen: An Automatic Benchmark Construction Toolkit for Evaluating LLMs on Multiple Sclerosis-MRI Imaging Knowledge
Abdul Basit, Muhammad Abdullah Hanif and Muhammad Shafique
Biomedical large
language model (LLM) evaluation requires auditable assessment of narrow, evolving, source-grounded subspecialty knowledge. Multiple sclerosis MRI (MS-MRI) provides a high-stakes textual-knowledge test case because correct
reasoning requires current diagnostic criteria, standardized acquisition and reporting knowledge, longitudinal monitoring concepts, lesion morphology, and recognition of difficult mimics. We present MS-Exam-Gen, a reproducible framework for constructing and auditing a text-based multiple-choice question (MCQ)
benchmark for MS-MRI knowledge; it does not evaluate direct MRI image interpretation. MS-Exam-Gen targets source-grounded criteria, protocols, reporting, and differential diagnosis. The framework combines expert-source indexing, exam-oriented topic induction, evidence-grounded MCQ generation, automated quality audits, a same-family consistency screen, and empirical calibration. From a 66-source corpus indexed into 4,289
retrieval chunks, the pipeline produced a locked 3,058-item candidate
benchmark spanning 16 topics and 53 subtopics. Evaluation across 12 primary LLM endpoints yielded 36,696 item-level predictions and separated performance over a 42.8-percentage-point accuracy range (89.7% to 46.9%). Across these endpoints, 25.5% of items were missed by at least four. Post-generation audits showed that refreshed construction reduced measurable answer cues, while option-order testing showed that absolute MCQ scores remain position-sensitive. Generated construction labels remain metadata rather than validated psychometric categories. Because expert adjudication and full option-order counterbalancing remain future work, MS-Exam-Gen is not a clinically certified examination. It should be interpreted as an automatically filtered, source-grounded candidate
benchmark and reproducible audit workflow for item-level and topic-specific LLM evaluation.
Read the full abstract ▾
Computation and Language
Introducing Code-Switched Contexts to Cognitively-Inspired Bilingual Model Training
Zhuojing Huang, Luise Pohlmann and Lisa Beinborn
During
language acquisition, bilingual children are regularly exposed to code-switched input and use it as a cognitive scaffold to accelerate vocabulary growth and cross-linguistic syntactic mapping. In contrast, computational bilingual models are conventionally pretrained on interleaved monolingual corpora. While introducing synthetic code-switching during pretraining has become a promising strategy to enhance cross-lingual
alignment and downstream performance, the structural and developmental parameters governing the success remain poorly understood. In this work, we investigate the efficiency of training with synthetic code-switched data across two typologically distinct
language pairs by controlling two key variables: the structural location of code-switches and the dynamic switching rate across training stages. Our results show that training with code-switched data improves cross-lingual
alignment for typologically close
languages.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.CV
Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts
Seungmin Oh, Seunghun Kang and Jongbin Ryu
Vision-
language models (VLMs) achieve strong zero-shot transferability but remain vulnerable to target-domain shifts at inference time. Test-time adaptation (TTA) offers a practical remedy, yet most existing VLM-TTA methods follow a prediction-side adaptation paradigm. They use test samples to adjust logits, prototypes, caches, priors, or feature statistics, often incurring additional computational overhead. In this paper, we take a different perspective and reframe VLM-TTA as candidate verification rather than prediction adjustment. We propose Test-Time Correction (TTC), a hypothesis-based correction framework guided by a simple principle: hypothesize, reconstruct, correct. Given a test feature and its top-k candidate labels, TTC treats each candidate label as a hypothesis, reconstructs the feature within the corresponding latent subspace stored in a
memory bank, and measures the resulting divergence shift. This shift quantifies how much the candidate subspace and its relations to other candidates change after the hypothetical insertion of the test feature. A correct candidate hypothesis induces only a small shift, whereas an incorrect one perturbs the subspace more strongly. TTC therefore corrects the prediction by selecting the candidate with the minimum aggregated divergence shift. This training-free candidate-verification mechanism avoids iterative optimization and provides a favorable accuracy-efficiency trade-off. Across five TTA settings and 15
benchmark datasets, including zero-shot classification, domain generalization, few-shot classification, base-to-novel generalization, and cross-
dataset evaluation, TTC consistently improves accuracy over state-of-the-art VLM-TTA methods while achieving up to 2x speedup, over 3x lower CPU
memory usage, and up to 1.4x lower GPU
memory usage than the lowest-
memory training-free baseline.
Read the full abstract ▾
Computation and Language
From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment Simulation
Quanyu Long, Xiao Chen, Jianda Chen et al. (+4 more)
Realistic environment replicas are increasingly valuable for training and evaluating LLM
agents, yet the original systems may be inaccessible or impractical to reproduce. We explore agentic
language world modeling: rather than rebuilding an executable environment, a world model
agent serves as the environment for a task
agent and supports faithful and stateful simulation. We instantiate this paradigm with Trace2Env, a learning-free framework for settings where the original system is unavailable but historical interaction traces remain accessible. Trace2Env reconstructs these traces into a reusable environment worldbook containing environment schemas, grounded evidence, and induced behavioral knowledge. At runtime, the world model
agent actively consults the worldbook together with persistent episodic state to infer each action's observation and lasting state effects. Across nine environments, Trace2Env improves both next-observation fidelity and long-horizon interaction consistency over conventional
prompt-based LWMs. In multi-turn interaction, task
agent actions generated against Trace2Env remain valid more often when replayed in the real environment, indicating that its simulated dynamics better preserve the consequences of earlier actions across successive turns. These results establish agentic
language world modeling as an alternative direction for building realistic environment replicas without reconstructing the original executable system.
Read the full abstract ▾
All authors Quanyu Long, Xiao Chen, Jianda Chen, Haozhen Zhang, Qisheng Hu, Jianzhu Bao, Wenya Wang
Computation and Language
Cross-Lingual Transferability of Training Data Extraction Attacks to Recover Memorized PII
Alexandru Nazare, Agnese Profico, Nicolò Vania et al. (+6 more)
The robustness of Personally Identifiable Information (PII) protection in Large
Language Models (LLMs) is a critical concern, yet the risks associated with cross-lingual data extraction remain under-explored. This study evaluates the vulnerability of English-centric and multilingual models to Training Data Extraction (TDE) attacks when prompted in non-English
languages. We construct a multi-domain PII
dataset comprising social media handles, email addresses, and phone numbers and translate the attack contexts into Italian, Spanish, French, and German. Our results show that TDE attacks against both English-centric and multilingual models transfer to different
languages: the attacks are successful on translated
prompts, even though only the original English
prompt might have been included in the pre-training data. A web-presence check on a sample of the translations confirms that they are not available online. The share of English leaks recovered in other
languages grows with the multilingual capability of the model, and it drops sharply when the original wording is lost, even without a change of
language. This suggests that native multilingual pre-training facilitates the emergence of latent cross-linguistic bridges that simplify the
retrieval of personally identifiable information (PII). We analyze the activations of multilingual large
language models (LLMs) and find that different translations of the same
prompt are bridged in similar representations, with the strongest
alignment in the middle layers. Our results highlight a fundamental security gap in modern LLMs, necessitating more robust,
language-agnostic sanitization strategies for future model
alignment.
Read the full abstract ▾
All authors Alexandru Nazare, Agnese Profico, Nicolò Vania, Elena Di Croce, Daria Caramanica, Davide Venditti, Elena Sofia Ruzzetti, Giancarlo A. Xompero, Fabio Massimo Zanzotto
Computation and Language · Cross-listed from cs.MA
Attention Tax, Handoff Tax: A Stylised Model of When Multi-Agent LLM Systems Help
Akshit Anchan and Nayonika Sen
Recent work on multi-
agent LLM systems reaches sharply different conclusions: some results show that a single
agent with the same information and compute should dominate a delegated system, others that multi-
agent gains grow with task depth. We argue that much of the disagreement comes from modelling different bottlenecks, and introduce a stylised reliability model built around two trade-offs. Decomposition reduces the burden of long contexts but incurs a handoff tax when information is compressed or transferred between
agents. Redundancy gains from multiple samples, but its benefit depends on how much their failures are shared. With
reasoning budget, verification, and task structure added, the model yields two crossover conditions: decomposition becomes preferable once the
attention cost avoided by resetting context exceeds the handoff cost, and parallel sampling at equal budget is eventually preferable when its shared-failure floor lies below the error floor of one
agent thinking longer. We connect these regimes to recent theoretical and empirical results. On a ledger-reconciliation task we measure the context-degradation curve and the handoff tax from single-
agent and handoff runs alone. From these the model places the crossover at depth 10 and predicts decomposition to win at depths 20, 50, and 100. It does, on step-level and final-balance accuracy, and the decomposed system's success, which the prediction never sees, lands within 9 percentage points of the predicted rate at every depth.
Read the full abstract ▾
Computation and Language
TrustMI: Causally controlling how assistants trust their users
Théo Lasnier, Romain Froger, Maxence Lasbordes et al. (+1 more)
Large
Language Model (LLM) assistants routinely decide whether they can trust users and third parties whose competence, intentions, and integrity they cannot verify. This uncertainty matters for safety, as trusting the wrong party can lead an
agent to comply with harmful requests or act on malicious instructions encountered during tool use. To study this problem, we define trust as an assistant's willingness to accept vulnerability to the actions of another party and ask whether such behavior can be causally controlled through model activations. We build 2,000 contrastive conversations spanning ability, benevolence, and integrity, where paired responses complete the same request but differ in whether the assistant trusts the user. From these pairs, we learn steering
matrices while keeping the model parameters frozen and test them across six instruction-tuned models from three families, finding that steering changes trust decisions monotonically in both directions. We then ask whether this effect extends to several safety-related
agent settings involving harmful requests,
prompt injections, and insider threats, while using benign-task and
reasoning as controls. Our findings provide evidence that trust in the user can be causally controlled along linear directions in model activations and provide a way to study how trust shapes safety-relevant behavior in
language models.
Read the full abstract ▾
All authors Théo Lasnier, Romain Froger, Maxence Lasbordes, Djamé Seddah
Computation and Language · Cross-listed from cs.CV
ROT: Rotating Hidden States towards Contextual Vectors for Hallucination Mitigation in LVLMs
Yijing Du, Xiangcheng Zhan and Shuo Yang
Large
Vision-
Language Models (LVLMs) frequently suffer from object
hallucination. Existing training-free interventions primarily manipulate
attention weights, which indirectly affect the deep semantics reaching the final predictive layers. In this work, we shift our focus to the hidden state vectors extracted after self-
attention and residual addition. Empirical analysis reveals that hallucinated
tokens do not simply over-rely on linguistic priors; instead, they exhibit an anomalous contextual deviation, showing significantly lower similarities to both textual and visual contexts in intermediate layers. Motivated by this, we propose ROT, a layer-specific, training-free framework. ROT dynamically detects semantic deviation in the middle layers and applies a norm-preserving rotation to steer the hidden states back toward the local
multimodal context plane spanned by the contexts. For subsequent layers, a representational smoothing mechanism is introduced to stabilize the calibrated trajectory. Extensive experiments on multiple
benchmarks demonstrate that ROT consistently reduces
hallucinations across various model architectures and scales, offering an efficient, geometry-driven solution for grounded generation.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.CR
Backdooring Sparse Autoencoders
Enrico Ahlers, Daniel Passon, Tobias Kiecker et al. (+2 more)
Sparse autoencoders (SAEs) are increasingly used not only to interpret
language models but also to intervene on their internal representations. We show that this creates a supply-chain attack surface: a maliciously modified SAE can induce attacker-chosen behavior when inserted into the forward pass of an otherwise unchanged
language model. We introduce a decoder-only SAE backdoor that leaves both the underlying LLM and the SAE encoder frozen, restricting the attack to a single auxiliary component at a single insertion layer. Using code generation as a case study, we demonstrate high rates of unsolicited code insertion across three
language models and a wide range of insertion layers, as well as trigger-dependent behavior conditioned on a
prompt cue. We further evaluate the modified SAEs using HumanEval and selected SAEBench metrics. While attack effectiveness varies across models and layers, strong backdoor behavior can coexist with relatively small changes in several conventional SAE quality measures. These results establish that SAEs can carry behavioral backdoors without modifying the
language model itself and should therefore be treated as security-sensitive components.
Read the full abstract ▾
All authors Enrico Ahlers, Daniel Passon, Tobias Kiecker, Eik Reichmann, Lars Grunske
Computation and Language
LightMTP: Lightweight Latent Multi-Token Prediction
Tamara Czinczoll, Julie Kallini, Gerard de Melo et al. (+1 more)
Next-
token prediction (NTP) is the standard pretraining objective for large
language models, yet it provides an explicit training signal only for the immediate next
token, which can lead models to exploit local patterns instead of capturing longer-range structure and ideas. Multi-
token prediction (MTP) addresses this by training models to predict several future
tokens. However, existing MTP methods often introduce a large number of new parameters with limited improvements in downstream performance. Latent MTP approaches address this efficiency issue by encoding future
tokens into a vector representation. However, these approaches usually rely on external helper models for future
token encoding. We propose LightMTP, a lightweight, i.e., parameter-efficient, latent MTP approach that bootstraps the future
token representations from the model's own hidden states. Our two LightMTP variants extend supervision to more future
tokens without requiring the additional computational overhead of conventional MTP nor the external supervision latent MTP normally relies on. LightMTP adds at most 1% extra parameters, retains better performance on general
language modeling
benchmarks, and achieves similar gains in planning, coding, and
reasoning.
Read the full abstract ▾
All authors Tamara Czinczoll, Julie Kallini, Gerard de Melo, Chen Shani
Computation and Language
Differentiable Bit-Widths: Co-optimizing Pruning and Quantization via SVD for Ultra-Efficient LLM Compression
Hankyul Kang and Jongbin Ryu
SVD-based
pruning and quantization have recently emerged as a promising strategy for the ultra-efficient compression of large
language models. In these methods, compression is performed in two stages: components are first truncated, and the remaining ones are subsequently quantized. Although this decoupled pipeline benefits from both
pruning and quantization, it requires separate optimization for each stage and fails to fully exploit their balance, which can lead to suboptimal performance under aggressive compression. To address this limitation, we propose a new LLM compression method that co-optimizes
pruning and quantization in a unified framework. Our key idea is a differentiable method for learning component-wise bit-widths, allowing less important components to be assigned 0-bit precision and pruned away. Notably, our method performs favorably against two-stage baselines, even when subjected to extreme quantization settings ($1.61$ bits) designed for ultra-efficiency. Code: https://github.com/MMAI-Laboratory/DBW.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.CV
Investigating Query-Insensitive Behavior in Spatio-Temporal Video Grounding
Eryk Kołodziejczyk, Alberto Presta, Karol Szurkowski et al. (+1 more)
Spatio-temporal
video grounding (STVG) aims to localize objects or events described by natural
language queries in both space and time. Existing STVG models are typically trained and evaluated under the assumption that each query is relevant to the input
video. In this work, we challenge this assumption by studying the behavior of state-of-the-art STVG models under irrelevant queries and missing textual input. Our experiments show that current models can still produce plausible spatio-temporal predictions even when the query is unrelated to the
video or removed entirely. We further analyze HCSTVG-v2 and VidSTG to identify
dataset regularities that may encourage such query-insensitive behavior. Our study highlights an underexplored limitation of STVG models and motivates negative-aware evaluation protocols and architectures that explicitly assess query relevance.
Read the full abstract ▾
All authors Eryk Kołodziejczyk, Alberto Presta, Karol Szurkowski, Michal Byra
Computation and Language
D-Loop: Looped Diffusion Drafting for Speculative Decoding
Kecheng Chen, Yuyang He, Cheng Gong et al. (+8 more)
Block
diffusion accelerates speculative decoding by drafting multiple
tokens in one forward pass. However, each position predicts a marginal distribution without observing earlier proposed
tokens, limiting draft quality and acceptance length. We identify a concrete failure, the repetition trap, in which neighboring positions produce redundant copies of the same
token. We explain this tendency theoretically and empirically examine its association with shorter accepted drafts. Recent methods refine marginal predictions with an additional causal head or a separately trained drafter, increasing parameter storage and introducing separate training objectives. We instead propose D-Loop, which introduces intra-block causal conditioning within the original
diffusion drafter without additional model components. Inspired by semi-autoregressive generation and parameter sharing, D-Loop reuses the same backbone across looped passes. The first pass proposes a block, and the second conditions on a selected prefix to regenerate the suffix in parallel. A complementary prefix–suffix objective trains the shared drafter for both anchor-only prefix prediction and prefix-conditioned suffix prediction. Across eight math, code, and chat
benchmarks, D-Loop can beat DFlash and DSpark on Qwen3-4B and Qwen3-8B with obvious gains.
Read the full abstract ▾
All authors Kecheng Chen, Yuyang He, Cheng Gong, Hui Liu, Guoping Long, Jiajun Li, Shi Wu, Suiyun Zhang, Haoliang Li, Ziru Liu, Rui Liu
Computation and Language
Breaking the Tie: A Cluster-Aware Routing Framework for Large Language Models
Yao Lu, Zhaiyuan Ji, Yaxin Gao et al. (+6 more)
With the rapid development of artificial intelligence, the emergence of various Large
Language Models (LLMs) has created a rich model ecosystem. However, this also brings a key challenge: how to select the optimal model for a specific user query. LLM routing addresses this need by dynamically assigning queries to the most suitable expert in the pool of candidate models. However, existing routing frameworks often simplify this process to a standard classification task; thus, a critical vulnerability is exposed when multiple candidate models correctly answer the same query. We formalize this capability overlap as routing noise, which misleads the router with arbitrarily correct candidate models, ultimately leading to routing collapse (a severe decline in generalization ability on unseen tasks). To address this problem, we propose a novel Cluster-Aware Soft-Labeling Routing (CASLR) framework. CASLR shifts the evaluation paradigm from the success of a single query to macro-domain consensus by replacing traditional one-hot vectors with a masked softmax mechanism. Specifically, for experts who answer incorrectly, we penalize their target probability to zero; for the remaining candidates, we directly compute continuous fine-grained soft labels based on their global clustering utility scores. We then use these refined soft labels to supervise a lightweight router. Specifically, the framework not only demonstrates superior accuracy on multiple
benchmarks, but also outperforms Llama-3.3-70B-Instruct by 7.80% in overall average performance. Furthermore, the extremely low routing inference latency of only 1.13s further confirms that CASLR can achieve efficient system scheduling with almost zero additional overhead, while ensuring high response quality.
Read the full abstract ▾
All authors Yao Lu, Zhaiyuan Ji, Yaxin Gao, Zeyu Wang, Zhe Tang, Jiaheng Wei, Zhaowei Zhu, Shanqing Yu, Qi Xuan
Computation and Language
Byte Language Models: Scaling, Emergent Abstractions, and Information Allocation
Jie Wang, Shiwei Luo, Qi Zhang et al. (+1 more)
Tokenizer-free
language models remove the inductive bias of fixed tokenizers by modeling text directly as bytes, but the resulting longer sequences substantially increase computation and eliminate explicit text abstractions. We ask whether this additional computation can be useful, and whether standard
Transformers can learn the abstractions that tokenization provides. We study these questions on
Transformers without specialized tokenization-related architectures. With
token-superposition training and hash
embeddings, byte
Transformers consistently outperform subword
Transformers as model size scales. We further find that byte
Transformers build local text abstractions as external tokenizers: a set of segmentation-like positions are used to collect local context representations, and restricting up to $25\%$ of intermediate layers to these local representations preserves downstream performance. Finally, these learned structures induce highly non-uniform generation difficulty, with uncertainty concentrated near local structure boundaries; exploiting them for speculative decoding yields $3.4\times$ more accepted
tokens than in subword
Transformers.
Read the full abstract ▾
All authors Jie Wang, Shiwei Luo, Qi Zhang, Yuanbin Wu
Computation and Language · Cross-listed from cs.LG
StagQ: Constraint-Driven Multi-Precision Weight Quantization for LLMs
Zhe Wei, Mengqi Guo, Yuan Yuan et al. (+3 more)
Serving a large
language model (LLM) across a fleet of deployments requires several weight-precision operating points. Multi-precision formats serve them all from one stream whose prefixes are valid lower-precision codes, instead of storing multiple copies. We present StagQ, a multi-precision weight format whose main stream is a 2-bit group-wise affine base followed by a configurable number of 1-bit refinement planes on a dyadic step schedule. Every supported precision is a readable prefix, decoded by an affine map derived from metadata shared across all precisions, with no per-weight lookup. A sparse side record, filled both before and after the grid is fitted, holds out the few weights the grid serves worst. We report two configurations of the encoder. At two bits the cheaper one leads the strongest multi-precision baseline on Llama-3.1-8B, Phi-4, and OLMo-2-7B by 3.1 to 7.0 MMLU points, at a slightly lower logical rate. At three bits it leads on Llama-3.1-8B, leads on Phi-4 at a higher rate, and ties on OLMo-2-7B. At four bits it ties on all three, at a higher rate. In a batch-one
matrix-vector product on an NVIDIA A100 GPU, timed on synthetic weights, our kernel is faster than the two baseline kernels in most shape-precision cases.
Read the full abstract ▾
All authors Zhe Wei, Mengqi Guo, Yuan Yuan, Jiunn Bin Lim, Boyi Pan, Michael Bi Mi
Computation and Language
Can Language Models Learn to Reject Their Own Bad Reasoning Steps?
Siheng Xiong, Xiaoze Liu, Yiqiao Jin et al. (+2 more)
Verifier-guided decoding can prevent harmful
reasoning steps from contaminating subsequent generation, but typically relies on an external learned verifier. We ask whether a
language model can instead reject its own bad
reasoning steps. We define a prefix's recoverability as the probability that the frozen generator can complete it correctly. Diagnostics show that adjacent recoverability changes are often difficult to resolve with practical Monte Carlo budgets, while same-prefix candidates exhibit a sparse low-recoverability tail. We introduce Self-Step Rejection (SSR), which trains a lightweight LoRA acceptance gate on the generator backbone while keeping the base model frozen. SSR uses confidence-qualified first-passage supervision: steps before the first resolved crossing of a root-relative recoverability barrier are accepted, the crossing step is rejected, and unresolved steps and suffixes are excluded. Training combines pointwise classification, same-prefix pairwise learning, and group-relative policy refinement using final-answer correctness. At inference, SSR accepts candidates or resamples from the unchanged prefix under rejection budgets, without an external learned verifier. Across three
reasoning models and five mathematical
reasoning benchmarks, SSR improves macro-average accuracy over single-pass decoding by 5.4–10.1 points using 1.21–1.40x as many generated
tokens, and achieves the highest macro-average accuracy among evaluated step-level methods. Full-solution scaling methods require 4.47–8.27x the single-pass
token cost for comparable performance.
Read the full abstract ▾
All authors Siheng Xiong, Xiaoze Liu, Yiqiao Jin, Xiaoqian Wang, Jing Gao
Computation and Language
HuatuoGPT-3: RL-Only Domain Adaptation from Base Models
Junying Chen, Xinyuan Xie, Ziniu Li et al. (+7 more)
Domain adaptation aims to turn a general-purpose large
language model (LLM) into an expert for a target domain. While the dominant SFT+RL pipeline offers a convenient cold start, it may reduce exploration diversity and introduces additional complexity through multi-stage optimization. These limitations motivate RL-only adaptation. However, pure on-policy RL suffers from a cold-start problem, while mixed-policy RL still falls short: informative
tokens in teacher outputs are learned too slowly in early training, and stale teacher outputs can hinder later improvement. We identify these two failure modes as Gradient Starvation and Teacher-Distribution Anchoring. To address them, we propose One-stage Policy Optimization (OnePO), which treats teacher outputs as transient guidance for policy improvement. OnePO combines Adaptive Objective
Evolution to strengthen learning on informative low-probability teacher
tokens and Teacher Retirement to discard teacher outputs once the current policy can surpass them. On medical adaptation, OnePO achieves 67.2 on HealthBench (Total) with only 20K training samples, outperforming SFT+RL and pure RL by 2.7 and 7.4 points, respectively. We further scale OnePO to produce HuatuoGPT-3, an open-source medical LLM series whose 27B variant reaches 70.1 on HealthBench (Total) and 71.4 on HealthBench Professional, surpassing frontier models such as GPT-6 Astra. Models and code are available at https://github.com/FreedomIntelligence/HuatuoGPT-3.
Read the full abstract ▾
All authors Junying Chen, Xinyuan Xie, Ziniu Li, Wenyuan Gu, Jianquan Li, Xiang Wan, Guangjun Yu, Ruoyu Sun, Haizhou Li, Benyou Wang
Computation and Language · Cross-listed from cs.CV
TasteRoute: Personalized Routing for Video Generation
Zhi Rui Tam, Chao-Chung Wu, Sin-Han Yang et al. (+8 more)
Rapid progress in
video generation has led to a plethora of models that differ substantially in capability and generation cost. This raises a natural question: can each request be efficiently routed to an appropriate model? We find that even when the consensus of the other annotators is used as an oracle, it agrees with each annotator's own favorite only 34-55% of the time. Motivated by this observation, we introduce TasteRoute, a personalized
video-generation router that selects a generator jointly based on the input request, user preferences, and available generation budget. Across text-to-
video and image-to-
video settings, TasteRoute is competitive with strong simple baselines on preference routing while reducing average generation cost. The cost saving increases under higher budget caps. Finally, we release TasteRoute-3k, a human-annotated
dataset containing multi-model
video comparisons, quality judgments, preference rankings, and user-profile signals to facilitate future research on personalized and cost-aware
video routing.
Read the full abstract ▾
All authors Zhi Rui Tam, Chao-Chung Wu, Sin-Han Yang, Peyton Ku, Brendan Kuang, Tzu-Ting Hsieh, Min-Fang Hsu, Fang-Ling Tsai, Yun-Nung Chen, Wei-Chiu Ma, Chieh-Yen Lin
Computation and Language
Noise Out, Bias In: Targeted Bias Injection in Diffusion Language Models via Closed-Loop Activation Steering
Sarim Hashmi, Mukul Ranjan, Abdelrahman Elsayed et al. (+3 more)
Masked
diffusion language models (dLLMs) generate text by iteratively denoising masked positions, re-predicting each
token multiple times before it is committed. An autoregressive decoder exposes an answer's distribution once, at the step that commits it; a dLLM exposes it at every denoising step before commitment, and we show that an adversary can exploit this. Since an answer remains open to revision over many denoising steps, an adversary with access to internal activations can watch how likely the model is to produce a chosen answer and adjust the intervention accordingly. Building on this observation, we study targeted bias injection, an attack that steers a frozen dLLM toward a demographic answer selected by the adversary. The attack uses a simple proportional-integral (PI) controller that tracks the target-answer probability during denoising and adapts the strength of a steering vector on the fly. On ambiguous BBQ questions where the correct answer is abstention, our attack raises LLaDA-8B-Instruct's preference for the targeted group from 1.8 to 16.7 percentage points, more than three times the strongest fixed-strength steering baseline, and on SocialStigmaQA it raises the selection of stigmatizing answers from 17.6% to 58.1%. Fitted to other demographic targets, the same attack shifts answers by up to 37 percentage points, and each attack takes about 40 minutes on one GPU. On the primary target, feedback is what makes the attack work: constant steering at the same average strength over the
token-committing steps produces a far smaller shift while corrupting nearly three times as many outputs, and a constant strength set separately for each example still falls well short. Our findings identify the denoising trajectory as a new control channel in dLLMs and call for bias audits that examine the serving stack rather than the frozen model alone.
Read the full abstract ▾
All authors Sarim Hashmi, Mukul Ranjan, Abdelrahman Elsayed, Muhammad Umer Sheikh, Fahad Shamshad, Nils Lukas
Computation and Language
Learning to Learn a Language
Lennart Carstens-Behrens and Holger Fröhlich
We present the Prior-Fitted
Language Model (PFLM), a 300M-parameter byte-level
transformer pretrained only on samples from a synthetic non-linguistic prior. Given a prefix of real text, it learns to predict the
language in context with frozen weights, having never seen a word of any real
language. Every training sequence is generated by a recurrent structural causal model drawn fresh from a distribution over such models. The model never sees the same
language twice during training, so the only way to predict the continuation is to infer the
language from the prefix. Samples from this prior share the statistical signatures of natural text: Zipfian frequencies, slow
entropy-rate convergence, and long-range dependence. On Wikipedia in six
languages, bits per byte fall from the uniform eight to between 0.9 and 2.4 at one million bytes of context. Given numerals instead of text, PFLM learns to count, to compare magnitudes, and to add approximately. It predicts deterministic sequences like Rudin-Shapiro or the
prime indicator, and it compresses six non-text domains, from source code to
speech, below gzip and PPMd. The model has not learned a
language. It has learned to learn one.
Read the full abstract ▾
Computation and Language · Cross-listed from cs.LG
Off-Policy Merging Beats On-Policy Self-Distillation for Continual Learning
Chen Henry Wu, Thomas Zhang and Aditi Raghunathan
A long-standing goal of AI is a model that can continually learn and improve itself. On post-trained models, supervised finetuning (SFT) on new data often causes poor generalization and catastrophic forgetting. As such, the conventional wisdom is that on-policy training is a prerequisite for continual learning. In practice, however, data containing new knowledge or capabilities are often off-policy. While methods such as on-policy self-
distillation (OPSD) try to bridge this gap by converting off-policy data into on-policy signal, they have been shown to cause
reasoning collapse. In this paper, we show that off-policy merging beats OPSD for continual learning. We first show that SFT learns a useful signal from new data, but naively applying its update interferes with existing capabilities. We reduce this interference with a simple recipe we term grafting, which changes where the update is learned and how it is applied: (1) learning the update on an earlier donor checkpoint, ideally even before the end of pretraining, and applying the weight update to the post-trained model; (2) scaling the weight update, equivalent to a form of model merging; and (3) optionally, masking the most sensitive update directions when the new data distribution is far from the post-trained model. Across continual learning settings including (1) distilling from expert traces, (2) self-improvement with
STaR and Pedagogical RL, and (3) injecting knowledge after pretraining cutoff, grafting Pareto-dominates both SFT and OPSD in new-task and old-task performance, while avoiding expensive on-policy sampling. Therefore, our work challenges on-policy training as a necessity for continual learning on RL-trained models.
Read the full abstract ▾
Computation and Language
HLA: Expressive Hybrid Linear Attention via Chunk-Wise Dynamic Mixing
Zhuokun Chen, Xi Lin, Xiyu Wu et al. (+3 more)
Linear
attention enables efficient long-context autoregressive decoding by compressing history into recurrent states, but this compression can make selective access to sparse and distant information difficult. Existing chunk-based extensions increase
memory capacity, yet learned chunk-mixing coefficients may remain fixed with respect to input content and therefore cannot adapt historical access to each query. We introduce Hybrid Linear
Attention (HLA), a query-dependent chunk-level
attention mechanism for Gated DeltaNet (GDN). HLA represents each completed chunk as an exact affine state transition and computes content-dependent routing gates from compact, self-attentively pooled representatives. Each gate interpolates the corresponding historical transition with the identity map, controlling both the chunk's additive
memory and its transformation of earlier states. Effective-support regularization further encourages concentrated routing for sparse inference. We evaluate HLA under both pretrained adaptation and from-scratch training. Across Qwen3.5 models from 0.8B to 9B, HLA consistently improves over native GDN and fixed chunk mixing, with gains of up to 5.57 percentage points on LongBench-V2 and 3.97 points on RULER. In a controlled from-scratch 1.3B setting trained for 100B
tokens with a 4K context, HLA also improves RULER performance from 4K to 32K, with gains increasing from 0.83 points at 4K to 4.22 points at 32K. These results demonstrate that query-dependent composition of recurrent
memory improves long-context modeling and remains effective beyond the training context while using compact per-chunk affine summaries. Project page: https://caesarhhh.github.io/hla/
Read the full abstract ▾
All authors Zhuokun Chen, Xi Lin, Xiyu Wu, Jiahao He, Jianfei Cai, Bohan Zhuang