arXiv_CL
-
CRM-Agent-Bench: Benchmarking Tool-Using Agents in Stateful Enterprise Environments
2026-10-23 00:05:21Monojit Banerjee, Zhujun Lan, Defu CaoAbstract
Deploying language agents in real applications demands more than the right answer: an agent must execute a multi-step business process correctly, follow domain-specific rules, and interact with a user across a dialogue, where the gap between saying and doing is the gap between a satisfied customer and a real loss. We argue that a deployable-agent benchmark should jointly assess four facets: the outcome an agent reaches, the trajectory it takes, its operational viability, and its policy compliance. Customer relationship management (CRM), where mutable records meet operational rules and consequential effects, exercises all four. We introduce CRM-Agent-Bench, an execution-grounded benchmark of 555 stateful, multi-turn tasks across nine business domains and 36 capability tags. Each task supplies a task-specific executable contract drawn from checks over actions, arguments, order, response, target state, and protected state; success is the strict conjunction of the checks declared for that task. We evaluate the complete agent system through these contracts and report scored coverage, latency, and cost. Across sixteen systems evaluated on the audited final release, strict success spans 64.32% to 83.60%; the closely grouped leading systems nonetheless diverge sharply in cost and per-domain reliability. CRM-Agent-Bench offers an auditable framework for measuring contract-grounded business-transaction integrity and adherence to task-declared benchmark policies, with a portable structure designed for extension as models improve. Github repository: this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2610.32004
PDF
-
DraMaze: A Multi-Agent Pipeline with Future-Aware Memory for Diverse Microdrama Script Generation
2026-10-22 23:54:48Siyang He, Qiqi Wang, Fan Liu, Yuncheng Huang, Pengfei Xu, Xi Chen, Ziwei He, Xipeng QiuAbstract
The rapid growth of microdramas has increased demand for high-quality script production. Long-form microdrama script generation with LLMs requires both diverse plot progression and long-range narrative coherence. Our observations reveal that candidate continuations increasingly concentrate on fewer narrative mechanisms as generation progresses, while conventional retrieval may reinforce trope reuse and overlook historical information essential to future plots. To address these challenges, we propose DraMaze, a multi-agent pipeline that explores diverse causal paths through trope memory-guided evolutionary tree search and mutation. It further constructs a Backward Narrative Dependency Tree to organize, retain, and retrieve historical memories based on future narrative goals. We also introduce MiDramaBench, a bilingual benchmark for evaluating script logic and creative quality across different script lengths, together with causal-pattern metrics for narrative repetition and diversity. Experiments show consistent improvements across three backbone LLMs. For 60-episode scripts of up to 60k words, DraMaze reduces causal-pattern repetition by 32.1% and achieves a Logic score of 82.5, outperforming the best-performing baseline by 8.5 points.
Abstract (translated)
URL
https://arxiv.org/abs/2610.32002
PDF
-
AusClinBench: a deterministically scored benchmark of large language models against current Australian clinical policy
Abstract
AI models already inform clinical decisions and draft notes in Australian clinical software, usually without the practitioner being told which model is used. Existing medical benchmarks are written for other jurisdictions. We wrote 142 questions from Australian guidelines and schedules across five domains (PBS authority, MBS billing, Ahpra obligations, triage red flags, immunisation and screening), each with one verifiable answer scored by exact match. Each question ran blind and with the governing clause supplied, which separates failure to recall the current rule from failure to apply it. We tested 28 models between July and October 2026. Blind accuracy ranged from 51.5% to 92.7%, and the highest-scoring models could not be separated. Of 791 failed model-question pairs, 684 (86%) were wrong blind and right with the clause. Confident-error rates ranged from 6.6% to 47.8%. On five immunisation questions where the guideline gives no directive, 82% of answers stated one. On 20 Aboriginal and Torres Strait Islander health questions, models did no worse overall, but on a few the wrong answers gave the rule for the general population, an older rule or another country's rule. We propose treating the model as an ingredient of a clinical product, with a minimum standard on local rules that sits beside product evaluation. The keys are one general practitioner's and are not yet independently reviewed. The benchmark measures base models, not clinical products, and does not license any model for clinical use.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31993
PDF
-
POS-Signature Demonstration Retrieval for In-Context Dependency Parsing
2026-10-22 22:15:48Muhammad Azam Afridi, Rajab Ali, Marco PassarottiAbstract
Demonstrations for in-context learning are often retrieved by semantic similarity, which may be a poor proxy for relevance in structured prediction. We study demonstration selection for Universal Dependencies parsing with a POS-signature policy that represents sentences with weighted part-of-speech statistics and retrieves three length-compatible, gold-parsed examples from the labeled training split. We evaluate two open-weight language models without fine-tuning on all 2,077 sentences of the UD English Web Treebank test set, comparing the policy with shot- and format-matched fixed and semantic three-shot baselines, as well as zero-shot, chain-of-thought, and critique-refinement systems. POS-signature retrieval yields the highest labeled attachment score among the tested prompting systems for both models (72.4 and 73.2), outperforming semantic retrieval by 1.9 and 2.7 points; paired-bootstrap 95% confidence intervals exclude zero for both contrasts. Chain-of-thought and refinement comparisons vary across models. These exploratory results favor the tested POS-signature retrieval approach on English-EWT. We release the prompts, archived outputs, and evaluation code.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31966
PDF
-
Which Answer Did the Seventeen-Year-Old Write? A Behavioral Case Study of Author-Age Misattribution, Miscalibrated Confidence, and Non-Corrective Self-Explanation in Six Frontier Language Models
Abstract
Three students answered the same exam prompt by hand: "What is happiness? Are you happy?" Their ages, recorded at collection, are 17, 28, and 45. The 17-year-old's answer is hedged and formally phrased. Asked to match each answer to an age and gender label, most frontier language models confidently assign it to an adult. This paper reports a structured case study of that failure across six models: Claude Fable 5, Gemini 3.1 Pro, GPT-5.6 Sol, Grok 4.5, GPT-4o (2024), and DeepSeek V4. We ran 1,081 API sessions and scored 3,225 responses across six experimental conditions that manipulate framing, disclosure format, and audience-awareness instruction. Three of six models perform at or near chance (1/3) on the matching task while reporting 73-88% mean confidence. The youngest writer's answer is misattributed to an adult in up to 90% of trials. Removing the multiple-choice scaffold collapses accuracy further: three models never place within three years of the true age once open-ended estimation is required. Institutional framing ("a community-college biology final") significantly degrades the best-performing model's accuracy; a neutral label does not. A single audience-awareness instruction significantly improves one model (+21 points) and moves another to ceiling, with no detectable effect on the three weakest models. Stated confidence is uncorrelated or inversely correlated with correctness for five of six models. After ground truth is revealed, one model names its own stereotyping in 68 of 114 explanations, yet shows the lowest accuracy and highest confidence of any model under blind presentation: articulate retrospective diagnosis that does not track the quality of its judgments. This is a three-item, single-topic, single-language case study meant to isolate a mechanism, not to estimate a population bias rate. Data, code, and raw model outputs are publicly released.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31946
PDF
-
Agreement rather than normalisation keeps depth reusable in looped language models
2026-10-22 21:42:16Ivan Viakhirev, Kirill Borodin, Maxim Abramov, Grach MkrtchianAbstract
Looped language models apply one block of layers repeatedly, and the design pays off only if depth is reusable: loop steps beyond the trained budget must remain usable at inference. We ask whether that rests on normalising the state between loop steps, as the released Ouro models do, or on agreement between the loop a model is trained with and the loop it is run with. On released weights this loop-boundary normalisation can be removed without adding parameters and with the readout path bit-identical, and the operation is an identity at one loop step, which serves as an implementation check. Removing it takes Ouro-1.4B and Ouro-2.6B to zero accuracy by the trained budget on a balanced binary task (chance 0.500). State-norm growth, the failure that prior accounts predict, is not the cause: the norm stays within a factor of 2.021 while the head commits to a single wrong token. Yet in 19M-parameter models trained from scratch on two synthetic tasks, 16 seeds per configuration, training without the normalisation is never worse than training with it and extrapolates better: the two are closest at the trained depth and separate on both sides of it (Cliff's delta up to 0.867 beyond the budget). Both results follow if agreement, and not the normalisation itself, is what keeps depth reusable. The damage sits in the readout together with the last layer of the block: in the models trained from scratch, refitting the head over the frozen modified loop recovers a third of the lost accuracy and plateaus, while refitting the last block layer with it recovers all of it. Finally, comparing looped with stacked depth at an equal inference loop count is confounded by unequal overshoot of the trained budget; at matched overshoot the advantage of looping vanishes on one task and shrinks on the other.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31943
PDF
-
LazyFlash: Forward-Dense, Backward-Sparse Attention for Accelerated Long-Context Fine-Tuning
2026-10-22 21:04:07Keqi Deng, Ruchao Fan, Jinyu LiAbstract
As large language models (LLMs) process increasingly long sequences, attention becomes a major computational bottleneck. Sparse attention mostly targets inference, yet many methods still train densely or cannot support sparse training, and their added complexity makes sparsity hard to turn into wall-clock speedups, leaving dense attention widely used in practice. This paper introduces LazyFlash, a forward-dense, backward-sparse attention method built on the observation that attention training is dominated by the backward pass. LazyFlash keeps the forward pass exactly dense, so the trained model remains a standard dense model that existing systems such as vLLM can serve without any change. For the backward pass, LazyFlash proposes a gradient-aware block sparsification design that ranks query blocks by their attention-output gradient norms and skips those contributing negligibly to parameter updates, turning the dense backward into a block-sparse one. A gradient-mass budget adapts the amount of skipped computation to each head and training step, while importance sampling preserves the dense gradient in expectation. These designs are implemented inside the tiled FlashAttention backward kernel, where skipped blocks are never loaded or computed, translating gradient sparsity into real wall-clock acceleration. Experiments on long-context text and speech LLMs show that LazyFlash maintains downstream performance while achieving up to 5.7x attention-backward speedup over FlashAttention and 1.7x end-to-end training speedup.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31911
PDF
-
TaxoSAE: Rethinking Sparse Autoencoders through Taxonomic Logic
2026-10-22 20:45:23Xinyu Guo, Minglai Yang, Clayton T. Morrison, Eduardo Blanco, Mihai SurdeanuAbstract
Recent Sparse Autoencoders (SAEs) decompose the representations of large language models (LLMs) into interpretable features, enabling precise readout and control of encoded knowledge. However, existing SAEs treat each feature as a standalone unit, whereas real-world knowledge is often taxonomically organized. As a result, readout carries no guarantee of taxonomic consistency, and control lacks the granularity to specify the scope of an edit. We propose TaxoSAE, which explicitly encodes taxonomical information, built on is-a (hypernym-hyponym) relations, into the SAE feature space. TaxoSAE mirrors a taxonomy with a stack of SAEs, one per taxonomy level, and introduces a taxonomic alignment loss that trains each concept to be active whenever one of its hyponyms is. On two factual QA taxonomies, one over biographical relations and one over animals, TaxoSAE enables (1) more accurate readout through the taxonomy with concentrated activations, e.g., +31.3 points in reading out the hypernyms of abstract concepts never queried in training; (2) more precise control, with +25.5 points in swap success, while suppressing a concept removes only its own answers (0-2% accuracy vs. 91-100% on other concepts); and, more importantly, (3) readout of unseen concepts through their hypernyms, with up to +37.9 points in hypernym readout accuracy.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31894
PDF
-
Mining Frames: Association Rule Mining as an Unsupervised Framework for News Framing Detection
2026-10-22 20:06:03Hager Hesham Mustafa (Anmat Computational Analysis Lab)Abstract
Computational framing analysis is dominated by supervised classifiers that require large annotated corpora and predefined frame taxonomies, limiting their use for non-English languages, novel events, and low-resource settings. This paper presents Association Rule Mining (ARM) as an unsupervised, annotation-free framework for news framing detection. Each document or sentence is treated as a transaction of lexical items, and high-lift co-occurrence rules are read as candidate framing signals (Gamson and Modigliani's condensing symbols), confirmed in context by journalist-researchers. The primary validation is Framing Gaza: 25,512 English-language articles from 16 US, UK, French, and German outlets covering the first year of the Gaza war (October 2023 to October 2024), analyzed with Apriori (mlxtend), platform-weighted support, POS-filtered n-grams, and lift-based scoring. ARM inductively surfaces interpretable patterns without predefined categories: a cross-platform military frame, {israeli, strike} to {kill}, appears as a closed rule in 14 of 16 outlets, and a month-by-month analysis recovers event-driven framing pivots (e.g., the April 2024 World Central Kitchen strike) with no prior event labeling. A systematic headline-body divergence is identified, with headlines carrying concentrated, high-lift editorial framing and body text reproducing institutional, procedural language. A methodological validation on the Seven Days project (1,575 Arabic articles on the 2026 Iran war) shows that FP-Growth enables sentence-level analysis with far lower memory overhead. Findings were validated through editorial uptake: two published investigative stories (Daraj Media; Saheeh Masr). Preprocessing code is released openly; the full pipeline is documented but withheld to preserve operational deployability.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31874
PDF
-
Rewriting Scaling Laws for Synthetic Data with Compute and Data Constraints
2026-10-22 19:56:45Nicholas Roberts, Ahmed Ahmed, Suhas Kotha, Percy Liang, Yu Wang, Alex Renda, Nitya ManiAbstract
Synthetic rewrites, where a language model rewrites a natural document e.g. into a target style, are now a standard ingredient of pretraining corpora. However, critical aspects of their use remain poorly understood, such as how much compute should be spent rewriting training data and whether a model trained on rewrites can outperform the rewriter. We study these two questions by proposing a rewriting scaling law that predicts loss from model size, training tokens, the fraction of training tokens that are rewrites, and the number of epochs over the natural data, charging the compute spent generating rewrites. We fit this scaling law in two settings. In the compute-limited setting, natural data is plentiful and rewrites come from a third-party rewriter trained with more compute than our largest run. We find that the estimated optimal rewrite fraction increases with compute, from under $10\%$ at $3\times10^{18}$ FLOPs to $32\%$ at $10^{22}$ FLOPs. Our scaling forecast extrapolates $33\times$ in FLOPs to train a 22B-parameter model, whose final loss is within $1.1\%$ of our prediction. This model achieves a better BPB than its rewriter on several web domains at one-tenth the rewriter's training compute. In the data-limited setting, we assume we only have a fixed seed set of natural data and that our rewriter is a model of the same size as the target, pretrained on that seed set. Again, we find the optimal rewrite fraction increases with the compute budget. We also demonstrate that a model trained on rewrites of its own pretraining data, generated by a rewriter of the same size, beats the best run on natural data alone at every model size up to 3B. Our results show that synthetic rewriting can be studied through the scaling law lens: that optimal methodological choices are predictable and scaling does not necessarily break down beyond the rewriter's capabilities.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31867
PDF
-
Aligning Large Language Models with Human Feedback: Mathematical Foundations and Algorithm Design
2026-10-22 19:50:27Siliang Zeng, Luca Viano, Chenliang Li, Jiaxiang Li, Volkan Cevher, Markus Wulfmeier, Stefano Ermon, Alfredo Garcia, Mingyi HongAbstract
This article provides an introduction to the mathematical foundations and algorithmic frameworks used to align Large Language Models (LLMs) with human intentions, preferences, and values. We discuss standard alignment techniques, such as fine-tuning (SFT), reinforcement learning with human feedback (RLHF), and direct preference optimization (DPO). We also explore the theoretical underpinnings of learning from human preferences, drawing connections to inverse reinforcement learning (IRL) and discrete choice models. We present state-of-the-art algorithms in a tutorial style, discuss their advantages and limitations, and offer insights into practical implementation. Our exposition is intended to serve as a comprehensive resource for researchers and practitioners, providing both a foundational understanding of alignment methodologies and a framework for developing more robust and scalable alignment techniques.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31862
PDF
-
Scaling cross-situational word learning on children's naturalistic experiences
2026-10-22 19:24:51Michael C. Frank, Alvin W. M. Tan, Khai Loong Aw, Robert Z. Sparks, Victor Moreno, Grace Keene-Lee, Meesha R. Ryan, Tarun Sepuri, Jane Yang, Virginia A. Marchman, Daniel L. K. Yamins, Bria LongAbstract
How do children learn the meanings of the simple, concrete words that form a majority of their early vocabulary? Statistical information about many word meanings can be recovered from the gradual accumulation of information about the co-occurrence of words with their referents, a proposal known as ``cross-situational learning''. Contrastive vision--language models allow this learning strategy to be tested on egocentric video data capturing children's early real-life experiences with visual and linguistic inputs. Initial successes of this method on small datasets provide a proof-of-concept, but vision models struggle to learn robust visual representations when trained on these same egocentric videos. How dependent is cross-situational word learning performance on the quality of the model's visual representations? And can cross-situational models scale to larger and more diverse datasets? To address these questions, we trained models on the BabyView corpus, including 2,441 hours of longitudinal egocentric video from 48 children. Models trained on this dataset learned object--label mappings, but performance was far higher when they used internet-pretrained visual representations as opposed to when visual representations were learned solely from the developmental video data. Further, the majority of the learning signal was carried by a small proportion of utterances that described a visible referent. Models learned lexical representations that showed recognizable human-like semantic structure, but model performance showed substantially different trajectories over training experiences than those observed in children. Cross-situational learning is a viable word learning strategy for early noun vocabulary, but constructing developmentally plausible models will require understanding how to learn high-quality visual representations from naturalistic data.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31842
PDF
-
SPECMEM: Accelerating Agentic Tool Calling via Live Memory Management
2026-10-22 19:24:21Zihe Song, Wenting Zhao, Ye Liu, Yuanjie Zhu, Ke Xu, Shanghao Li, Philip S. YuAbstract
Tool-calling agents are mostly optimized for accuracy; in deployment, however, their growing cost is latency: every tool call is decoded autoregressively. Speculative decoding accelerates this without changing the output, and retrieved past calls make strong drafts under real use, where each user's behavior tends to repeat. Yet existing systems keep no cross-session memory, or build the draft store once and never update it. We propose SpecMem, a tool-call memory kept alive and managed throughout deployment: verified calls are written back online into bounded per-user stores with LRU eviction, and a side-effect gate defers irreversible calls until verification confirms. On multi-session, multi-user query streams served by three open-weight models, a frozen store steadily loses acceptance while SpecMem holds it: a 27% gap in mean accepted tokens (+11% to +34% across benchmarks, tracking how much novel work arrives after the store is built) and up to 1.74x wall-clock speedup. Ablations attribute the gain almost entirely to keeping the store live; local per-user scoping adds no acceptance but closes a measured cross-user leakage channel, and gating on declared side effects, rather than on retrieval confidence, is what eliminates irreversible mistakes.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31840
PDF
-
Latent Coherence: Do Frontier Language Models' Probability Judgments Obey the Law of Total Probability?
Abstract
LawZero's Scientist AI trains a Predictor toward the Bayesian posterior, with an objective that penalizes inconsistent probabilities among latent explanations. We ask whether current frontier language models, with no such training, are already consistent in this sense zero-shot: do their probability judgments satisfy the law of total probability, $P(X)=\sum_z P(z)P(X|z)$, when the latent $z$ must be inferred from records in the context? In separate single-turn calls, models are asked for $P(X)$, for $P(z)$ over the latent's states, and for $P(X|z)$; a causal family applies the same test to $P(y|do(x))$ and the back-door adjustment. Every query has an exact ground-truth probability, and a subset of items is asked in three wordings to separate persistent incoherence from rewording noise. Across seven models, GPT-6.1 Sol and Claude Opus 5.5 are the most coherent (mean absolute residual 0.08 and 0.12 log-odds, about 2 percentage points). Claude Haiku 5.5, Gemini 3.8 Flash and DeepSeek V4-Pro show persistent incoherence (0.27-0.35 log-odds) that is mostly item-specific rather than a shared offset. Within every model, larger residuals go with larger errors against the truth (Spearman 0.40-0.62), so the residual, which needs no ground truth, flags unreliable answers. Reasoning length does not predict coherence, across models or within a model across rewordings of the same question. Simple consistency checks pass almost everywhere; the failures are specific to reasoning through a latent variable. Code and data: this https URL
Abstract (translated)
URL
https://arxiv.org/abs/2610.31833
PDF
-
Free Lunch for LLMs Confidence: The RL Critic Already Knows
2026-10-22 18:57:52Shujin Wu, Heng JiAbstract
Reliable confidence estimation is essential for trustworthy deployment of large language models (LLMs), but dedicated confidence training stage requires additional data and compute. We ask a fundamental question: what prediction problem does confidence training actually solve? Inspired by prior work on external confidence estimation, we compare self- and external estimators of comparable capacity trained on the same (question, response, correctness) triples. We show that an external model trained solely on a policy's behavioral traces can estimate the policy's confidence as accurately as the policy itself, without accessing the policy's internal representations. This points to a deeper connection with value function learning in reinforcement learning (RL): under a correctness-based outcome, we prove that confidence estimation and critic training are equivalent, both predicting the likelihood that a policy's response will be correct from its trajectory. A simple consequence follows: for LLMs trained with value-based RL, the existing critic can be reused as an off-the-shelf confidence estimator with no additional training or data. The policy remains unchanged, so its task performance is preserved, and the critic provides confidence throughout generation, including intermediate reasoning steps. Across math, code, tool calling, and general reasoning spanning a dozen domains, the reused critic is the best confidence estimator on every in-distribution split and the best-calibrated on 14 of 16, without hurting the task performance.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31820
PDF
-
AfriGuard: Safety and Robust Refusal For African Languages
2026-10-22 18:57:19Israel Abebe Azime, Jesujoba O. Alabi, Iffat Maab, Alexander D. Cai, Tadesse Destaw Belay, Atnafu Lambebo Tonja, Jonas Kgomo, Folasade Alabi, Salomey Osei, Saminu Muhammad Aliyu, Nkechinyere Faith Aguobi, Bontu Fufa Balcha, Blessing Sibanda, Davis David, Mouhamadane Mboup, Daud Abolade, Neo Putini, Louis Powell, Kiant'e Brantley, Philipp Slusallek, Dietrich Klakow, Julia KreutzerAbstract
Large language models (LLMs) designed to interact with users are typically trained to respond appropriately to harmful or toxic content. However, ensuring appropriate responses becomes more challenging in diverse linguistic and cultural contexts, where safety risks may be expressed or interpreted differently. Building safe AI systems therefore requires detecting safety risks, understanding their nature, and responding appropriately, while accounting for linguistic and cultural variations that add complexity to each stage. In this work, we introduce AfriGuard, a culturally grounded multilingual dataset for studying harmful intent, safety-category understanding, and refusal behavior across 10 African languages and English. Our findings reveal that guardrail models underperform both open- and closed-source LLMs on safety-intent and safety-category classification, while models show limited understanding of appropriate safety behavior, scoring below 60% F1. We further find that models often fail to respond appropriately to unsafe prompts not because they explicitly refuse, but because they produce linguistically inadequate responses that result in incidental non-assistance rather than proper refusals. Finally, using AfriGuard, we develop efficient guardrail models that achieve performance competitive with models up to 30x larger, introduce refusal detectors that improve average performance by 43.9% over existing detectors, and demonstrate that supervised fine-tuning can substantially improve overall safety performance.
Abstract (translated)
URL
https://arxiv.org/abs/2610.26474
PDF
-
ASTOR: Benchmarking and Improving Autoformulation for Soft-Constraint Optimization with a Formal Grammar
2026-10-22 18:21:47Amer Goel, Yeganeh Alimohammadi, Andrew DawAbstract
Real-world decision problems combine two qualitatively different kinds of constraints: requirements that must be satisfied, called hard constraints, and preferences that can be violated, called soft constraints. Because these problems are often posed to decision-makers in natural language, large language models offer a natural tool for assisting with their solution. This has given rise to autoformulation, which focuses on using LLMs to translate natural-language descriptions of decision problems into executable mathematical optimization models. However, this research often neglects soft constraints like those proposed by consumers, managers, and planners, despite the fact that many practical problems require modeling them explicitly. We introduce ASTOR, an inference-time framework that maps a natural-language decision problem to a grammar-guided abstract syntax tree (AST) before deterministically compiling the resulting representation into executable optimization code. We also introduce SoftOR, a unified benchmark of natural language problems which focuses on soft constraint formulations. On SoftOR, ASTOR achieves the best aggregate accuracy among comparable single-request and multi-agent LLM methods while remaining competitive on standard autoformulation benchmarks that include only hard constraints. Code and data are available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2610.26819
PDF
-
Direct Pass@$k$ Optimization for Thought Pruning: Balancing Quality and Diversity in Test-Time Reasoning
2026-10-22 17:59:48Mahdi Salmani, Adel Javanmard, Devansh Gupta, Meisam RazaviyaynAbstract
Test-time scaling has emerged as a key inference strategy for reasoning models, where a model generates a large number of candidate solutions in the hope of finding at least one correct solution. Running all candidate solutions to completion can be computationally demanding, so recent work prunes the candidate pool after generating only a short prefix, committing further computation only to the surviving candidates. Existing pruning strategies divide broadly into two categories: Quality-based methods assign a score to each partial solution and discard those predicted to ultimately lead to an incorrect final solution, while diversity-based methods select partial solutions that follow different reasoning trajectories and are therefore likely to cover different final solutions. The tension between these two objectives is an instance of exploration versus exploitation, yet little to no prior work on pruning exposes this trade-off directly, let alone optimizes it. In this paper, we introduce a pruning procedure that addresses this trade-off directly by optimizing pass@$k$ over the selected solution pool. In particular, it chooses the pool to maximize the probability that at least one of its partial solutions, when continued to completion, yields a correct answer. Because this objective simultaneously rewards high-quality partial solutions and penalizes redundancy among them, quality and diversity are naturally balanced by the criterion itself rather than through separate ad-hoc design choices. On eight open-weight reasoning models, our method outperforms existing quality-based and diversity-based pruning strategies. The gains are particularly pronounced when the pruning procedure keeps a moderate number of partial solutions for continued generation, highlighting the value of directly balancing quality and diversity during pruning.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31667
PDF
-
Cost-Aware Abstention for LLM-Based IT Ticket Classification
2026-10-22 17:58:52Sachin GuptaAbstract
Wrong automatic routing of an IT support ticket can cost more than human triage. We evaluate cost-aware abstention as an offline, cost-conditional readiness assessment: select a routing policy on calibration data under explicit costs, then evaluate that frozen policy once on held-out test data. On a public 47,837-ticket, eight-class corpus, we compare four trained non-LLM classifiers with four large language models under zero- and few-shot prompting. Under the illustrative cost model (c_ok, c_triage, c_wrong) = (0, 1, 5), TF-IDF with logistic regression and a threshold of 0.60 selected on calibration reduces test cost from 0.763 to 0.472, a 38.1% relative reduction (paired ratio-bootstrap 95% CI [35.7%, 40.4%]), while lowering wrong routes from 15.3% to 3.5%. Frozen calibration-selected policies also beat both always-route and always-defer on held-out test data in 14 of 16 settings in a post-hoc cost-parameter analysis. The strongest prompted configuration reaches 0.477 accuracy on a stratified 1,000-ticket test subset, versus 0.879 for fine-tuned DistilBERT on the same tickets. Isotonic lowers expected calibration error without reliably improving discrimination: GPT-4o-mini few-shot has one confidence value before and after recalibration and correctness AUROC 0.500. We therefore report correctness AUROC and tie-aware risk-coverage area alongside ECE. At the analytically derived break-even probability 0.80, seven prompted configurations defer all tickets; Haiku few-shot routes only 19/1,000. An LLM-only agreement control and margin/entropy rejection baselines further separate calibration from discrimination. The findings are scoped to this dataset, prompt design, model endpoints, and illustrative costs; the work is an evaluation framework, not a new abstention algorithm or universal deployment certification.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31657
PDF
-
Low-Bit N-Gram Embeddings for Language Models
2026-10-22 17:56:13Stephen Chung, Wenyu DuAbstract
Many recent open-weight language models use n-gram embedding tables to improve performance. These tables store learned representations of short token sequences, such as pairs or triples of consecutive tokens, which the model retrieves during processing. However, they can require substantial memory: DeepSeek-V4.1-Flash, for example, allocates 196 billion parameters to its Engram n-gram memory. We therefore investigate whether this memory cost can be reduced while preserving the performance improvement that the tables provide. We propose storing n-gram embeddings at low precision and systematically investigate different quantization methods and bit allocations across tables. Across nanochat-, Llama-, and GPT-2-family models trained from scratch, four-bit tables save approximately 75% of table memory while preserving more than 99% of the validation-loss improvement from adding the tables. Using fewer bits leads to a larger drop in performance, although we find that a short fine-tuning phase recovers a significant part of this loss. In addition, reinvesting the saved memory in larger tables can improve performance beyond that of the original full-precision tables while still using less memory. These table memory savings also extend to large existing models: on DeepSeek-V4.1-Flash and Qwen3.8-Flash-Next, low-bit tables save 47-55% of table memory, with perplexity increases of less than 1.2% and largely preserved downstream task performance. These findings establish low-bit precision as a practical way to reduce the memory requirements of n-gram embeddings and use the saved capacity to improve model performance.
Abstract (translated)
URL
https://arxiv.org/abs/2610.31646
PDF
