Searcharxiv⌕ Search

arXiv subjects

Pang Wei Koh

Publications and source records attributed to Pang Wei Koh.

At least 19 recordsLinked to original sources
arXiv2026

Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall

Logit-based knowledge distillation (KD) is used to train smaller language models (LMs) via supervision from stronger teachers, but whether its benefits are consistent across training stages remains unclear. Through controlled experiments, we find that forward Kullback-Leibler (KL) distillation--the standard KD formulation--with post-trained teachers behaves fundamentally differently during mid-training, an intermediate phase of self-supervised learning on curated corpora. Surprisingly, while forward KD simultaneously improves reasoning and factual recall during pre-training relative to standard next-token prediction (NTP), it instead slows factual recall acquisition during mid-training despite continued reasoning gains. We trace this stage dependence to an asymmetry in teacher confidence across data domains and the student's evolving knowledge state: teachers are more confident on procedural than knowledge-intensive data, while students acquire low-entropy factual knowledge earlier in training. To mitigate this imbalance, we propose Switch Distillation, a simple mid-training objective that distills on tokens where the teacher is confident, using teacher predictive entropy as a lightweight routing signal, and otherwise falls back to cross-entropy. Switch Distillation consistently outperforms existing distillation objectives across teacher sizes. Relative to standard NTP, it achieves 1.61-1.71x the reasoning performance and 1.13-1.19x the knowledge and commonsense performance while preserving 96.7-96.8% of factual recall. Crucially, these benefits persist after post-training: Switch Distillation closes the factual recall gap while maintaining 1.25-1.32x and 1.13-1.20x gains in reasoning and knowledge and commonsense, respectively.

cs.CL↗
arXiv2026

Efficient Test-Time Adaptation through Human-AI Interaction

AI agents are trained on population-scale data to encode broad capabilities spanning those of many practitioners. Yet the artifacts they produce rarely meet the personal bar professionals need to stake their reputation on. On realistic, open-ended tasks where success criteria are heterogeneous and insufficiently documented, individual expertise lives precisely in the elevation and departure from the average. In practice, iterative human-agent interaction surfaces criteria that users cannot fully specify up front, yet apply repeatedly across tasks. We argue this cross-session interaction data is a rich, underused signal for closing the gap to individual expertise. In this work, we propose test-time adaptation through human-agent interaction (TAHI), which integrates these signals into agent context and weights, and crystallizes each user's training and evaluation criteria via an evolving rubric module. We adapt agents to 30 individuals in two high-utility domains, writing and visual creation, on a total of 600 tasks. Our agents improve solo task success by 4.5-20.9% within only tens of tasks. Meanwhile, our evolving rubric module serves as a scalable annotation tool, creating evaluation rubrics that catch 16.0-22.3% more failures than those from LMs or humans alone. While agents are adapted towards individuals, we show these personalized agents also produce improvements in success of up to 8.8% that generalize across users.

cs.AI↗
arXiv2026

Anchored Decoding: Provably Reducing Copyright Risk for Any Language Model

Language models (LMs) tend to memorize portions of their training data and emit verbatim spans. When the underlying sources are sensitive or copyright-protected, such reproduction raises issues of consent and compensation for creators and compliance risks for developers. We propose Anchored Decoding, a plug-and-play inference-time method for suppressing verbatim copying: it enables decoding from any risky LM trained on mixed-license data by keeping generation in bounded proximity to a permissively trained safe LM. Anchored Decoding adaptively allocates a user-chosen information budget over the generation trajectory and enforces per-step constraints that yield a sequence-level guarantee, enabling a tunable risk-utility trade-off. To make Anchored Decoding practically useful, we introduce a new permissively trained safe model (TinyComma 1.8B), as well as Anchored$_{\mathrm{Byte}}$ Decoding, a byte-level variant of our method that enables cross-vocabulary fusion via the ByteSampler framework (Hayase et al., 2025). Across six model pairs on long-form metrics for copying risk and utility, Anchored and Anchored$_{\mathrm{Byte}}$ Decoding define a new Pareto frontier, preserving near-original fluency and factuality while closing up to 75% of the measurable copying gap between the risky baseline and a safe reference, at a modest inference overhead.

cs.CL↗
arXiv2026

Spokes: Optimizing for Diverse Pretraining Data Selection

Diversity plays a critical role in data selection, improving performance under fixed data budgets by reducing redundancy and repetition. However, optimizing for diversity is inherently challenging, as it is a set-level property that depends on interactions between data points rather than individual examples. As a result, existing approaches typically rely on proxies or approximations, which often fail to ensure sufficiently diverse subsets. In this work, we directly optimize diversity by introducing a probabilistic diversification framework based on the G-Vendi score, optimized via exponentiated gradient descent. Our method produces subsets that are substantially more diverse than those obtained via random sampling, achieving a +489 increase in G-Vendi score on a 500k-sample subset. We evaluate our approach on FineWeb and DCLM, where it consistently outperforms existing methods. Notably, SPOKES (diversity-only) improves average downstream performance by +0.4 and +0.5 points over random sampling on DCLM and FineWeb, respectively. More importantly, jointly optimizing for both quality and diversity yields the strongest results: SPOKES achieves gains of +1.5 and +1.4 points on DCLM and FineWeb, outperforming all baselines, including semantic deduplication and quality filtering.

cs.CL↗
arXiv2026

RLVE: Scaling Up Reinforcement Learning for Language Models with Adaptive Verifiable Environments

We introduce Reinforcement Learning (RL) with Adaptive Verifiable Environments (RLVE), an approach using verifiable environments that procedurally generate problems and provide algorithmically verifiable rewards, to scale up RL for language models (LMs). RLVE enables each verifiable environment to dynamically adapt its problem difficulty distribution to the policy model's capabilities as training progresses. In contrast, static data distributions often lead to vanishing learning signals when problems are either too easy or too hard for the policy. To implement RLVE, we create RLVE-Gym, a large-scale suite of 400 verifiable environments carefully developed through manual environment engineering. Using RLVE-Gym, we show that environment scaling, i.e., expanding the collection of training environments, consistently improves generalizable reasoning capabilities. RLVE with joint training across all 400 environments in RLVE-Gym yields a 3.37% absolute average improvement across six reasoning benchmarks, starting from one of the strongest 1.5B reasoning LMs. By comparison, continuing this LM's original RL training yields only a 0.49% average absolute gain despite using over 3x more compute. We release our code publicly.

cs.CL↗
arXiv2026

JobBench: Aligning Agent Work With Human Will

Current benchmarks for occupational AI agents are scoped primarily by economic values, telling a replacement story. We introduce JobBench, which evaluates AI agents on the workflows that experts identify as high-priority for delegation, empowering humans based on their needs instead of replacing them with GDP value. JobBench covers 130 agentic tasks across 35 occupations. Each task is packaged as a workspace of heterogeneous reference files, requiring the agent to reason through the cluttered information streams of real professional work. Outputs are graded by a fact-anchored chain of rubrics, averaging 35.6 binary criteria per task. We evaluate 36 models; the strongest, Claude Opus~4.7 under Claude Code, reaches only 45.9 %. We hope JobBench shifts the community's target labour-market effect from replacement to enhancement: building agents that do what humans actually want delegated, not only what is most economically valuable.

cs.AI↗
arXiv2026

DR Tulu: Reinforcement Learning with Evolving Rubrics for Deep Research

Deep research agents perform multi-step research to produce long-form, well-attributed answers. However, most open deep research agents are trained on easily verifiable short-form QA tasks via reinforcement learning with verifiable rewards, which does not extend to realistic long-form tasks. We address this with Reinforcement Learning with Evolving Rubrics (RLER), where rubrics are constructed and maintained to co-evolve with the policy model during training. This allows the rubrics to incorporate newly explored information from search and contrasting model responses, enabling better fact checking and more discriminative on-policy feedback. Using RLER, we develop Deep Research Tulu (DR Tulu-8B), the first fully open model that is directly trained for open-ended, long-form deep research. Across four long-form deep research benchmarks in science, healthcare, and general domains, DR Tulu substantially outperforms existing open deep research agents (by 15.6% over Tongyi DR on average) and matches or exceeds proprietary deep research agents (by 0.7% over OpenAI DR on average), while being significantly smaller and cheaper per query (1000x cheaper than OpenAI DR per query).

cs.CL↗
arXiv2026

Deep Reasoning in General Purpose Agents via Structured Meta-Cognition

Humans intuitively solve complex problems by flexibly shifting among reasoning modes: they plan, execute, revise intermediate goals, resolve ambiguity through associative judgment, and apply formal procedures to well-specified subproblems. Current LLM agents lack this flexibility, as their scaffolds hard-code such reasoning decisions in advance. These scaffolds are effective when their prescribed structure matches the task, but brittle when solving the task requires adapting the structure of reasoning itself. We introduce Deep Reasoning -- an inference-time approach for constructing task-specific scaffolds through structured meta-reasoning. Deep Reasoning uses a formal language that represents meta-reasoning as executable decompositions over associative inference, formal computation, and recursive subproblem solving, enabling decomposition principles to be encoded as in-context examples that guide test-time scaffold construction. We instantiate this approach in a general-purpose agent (DOLORES) that distributes complex tasks across more controlled reasoning threads. We evaluate it against state-of-the-art scaffolding methods across four hard benchmarks: multi-hop reasoning, long-chain question answering, long-context aggregation, and deep research-style information seeking. DOLORES outperforms all evaluated scaffolds across three model sizes and two model families, improving over the strongest evaluated scaffold baseline by 24.8% on average. DOLORES distributes cognition across structured, lower-load reasoning threads, thereby reducing premature termination and hallucinations. This advantage can even bridge the scaling gap, with an 8B version surpassing all evaluated 32B baselines from the same family in more than half the settings. These results point toward future agentic systems that treat scaffolding as adaptive reasoning, constructing the structure each task requires just-in-time.

cs.CL↗
arXiv2026

EvoLM: Self-Evolving Language Models through Co-Evolved Discriminative Rubrics

Language models encode substantial evaluative knowledge from pretraining, yet current post-training methods rely on external supervision (human annotations, proprietary models, or scalar reward models) to produce reward signals. Each imposes a ceiling. Human judgment cannot supervise capabilities beyond its own, proprietary APIs create dependencies, and verifiable rewards cover only domains with ground-truth answers. Self-improvement from a model's own evaluative capacity is a reward source that scales with the model itself, yet remains largely untapped by current methods. We introduce EVOLM, a post-training method that structures this capacity into explicit discriminative rubrics and uses them as training signal. EVOLM trains two capabilities within a single language model in alternation: (1) a rubric generator producing instance-specific evaluation criteria optimized for discriminative utility, which maximizes a small frozen judge's ability to distinguish preferred from dispreferred responses; and (2) a policy trained using those rubric-conditioned scores as reward. All preference signals are constructed from the policy's own outputs via temporal contrast with earlier checkpoints, requiring no human annotation or external supervision. EVOLM trains a Qwen3-8B model to generate rubrics that outperform GPT-4.1 on RewardBench-2 by 25.7%. The co-trained policy achieves 69.3% average on the OLMo3-Adapt suite, outperforming policies trained with GPT-4.1 prompted rubrics by 3.9% and with the state-of-the-art 8B reward model SkyWork-RM by 16%. Overall, EVOLM demonstrates that structuring a model's evaluative capacity into co-evolving discriminative rubrics enables self-improvement without external supervision.

cs.AI↗
arXiv2026

Olmo 3

We introduce Olmo 3, a family of state-of-the-art, fully-open language models at the 7B and 32B parameter scales. Olmo 3 model construction targets long-context reasoning, function calling, coding, instruction following, general chat, and knowledge recall. This release includes the entire model flow, i.e., the full lifecycle of the family of models, including every stage, checkpoint, data point, and dependency used to build it. Our flagship model, Olmo 3 Think 32B, is the strongest fully-open thinking model released to-date.

cs.CL↗
arXiv2026

Modeling Multi-Objective Tradeoffs with Monotonic Utility Functions

Countless science and engineering applications in multi-objective optimization (MOO) necessitate that decision-makers (DMs) select a Pareto-optimal (PO) solution which aligns with their preferences. Evaluating individual solutions is often expensive, and the high-dimensional trade-off space makes exhaustive exploration of the full Pareto frontier (PF) infeasible. We introduce a novel, principled two-step process for obtaining a compact set of PO points that aligns with user preferences, which are specified a priori as general monotonic utility functions (MFs). Our process (1) densely samples the user's region of interest on the PF, then (2) sparsifies the results into a small, diverse set for the DM. We instantiate this framework with soft-hard functions (SHFs), an intuitive class of MFs that operationalizes the common expert heuristic of imposing soft and hard bounds. We provide extensive empirical validation of our framework instantiated with SHFs on diverse domains, including brachytherapy, engineering design, and large language models. For brachytherapy, our approach returns a compact set of points with over 3% greater SHF-defined utility than the next best approach. Among the other domains, our approach consistently leads in utility, as a final compact set of just 5 points captures over 99% of the utility offered by the entire dense set.

cs.LG↗
arXiv2026

A False Sense of Privacy: Evaluating Textual Data Sanitization Beyond Surface-level Privacy Leakage

Sanitizing sensitive text data typically involves removing personally identifiable information (PII) or generating synthetic data under the assumption that these methods adequately protect privacy; however, their effectiveness is often only assessed by measuring the leakage of explicit identifiers but ignoring nuanced textual markers that can lead to re-identification. We challenge the above illusion of privacy by proposing a new framework that evaluates re-identification attacks to quantify individual privacy risks upon data release. Our approach shows that seemingly innocuous auxiliary information -- such as routine social activities -- can be used to infer sensitive attributes like age or substance use history from sanitized data. For instance, we demonstrate that Azure's commercial PII removal tool fails to protect 74\% of information in the MedQA dataset. Although differential privacy mitigates these risks to some extent, it significantly reduces the utility of the sanitized text for downstream tasks. Our findings indicate that current sanitization techniques offer a \textit{false sense of privacy}, highlighting the need for more robust methods that protect against semantic-level information leakage.

cs.CR↗
arXiv2026

PrefDisco: Benchmarking Proactive Personalized Reasoning

Current large language model (LLM) development treats task-solving and preference-alignment as separate challenges, optimizing first for objective correctness, then for alignment to aggregated human preferences. This paradigm fails in human-facing applications where solving a problem correctly is insufficient if the response mismatches the user's needs. This challenge intensifies in just-in-time scenarios where no prior user interaction history exists due to cold-start conditions or privacy constraints. LLMs need to proactively identify what they don't know about the user, strategically elicit preference values through questioning, then adapt their reasoning processes and responses accordingly -- a complicated chain of cognitive processes which we term personalized reasoning. We introduce PrefDisco, an evaluation methodology that transforms static benchmarks into interactive personalization tasks using psychologically-grounded personas with sparse, context-dependent preferences, and define PrefAlign as a fine-grained rubric-based metric for measuring preference alignment. PrefDisco builds scenarios where identical questions require different reasoning chains depending on user context, as optimal explanation approaches vary by individual expertise and preferences while maintaining factual accuracy. Evaluation of 21 frontier models across 10 tasks reveals 29.0% of naive personalization attempts produce worse preference alignment than generic responses, yet generic responses also fail to serve individual user needs. These findings suggest personalized reasoning requires dedicated development rather than emerging naturally. PrefDisco provides a foundation for developing systems that can adapt to individual users in education, healthcare, and technical domains where personalization is critical.

cs.CL↗
arXiv2026

PanCanBench: A Comprehensive Benchmark for Evaluating Large Language Models in Pancreatic Oncology

Large language models (LLMs) have achieved expert-level performance on standardized examinations, yet multiple-choice accuracy poorly reflects real-world clinical utility and safety. As patients and clinicians increasingly use LLMs for guidance on complex conditions such as pancreatic cancer, evaluation must extend beyond general medical knowledge. Existing frameworks, such as HealthBench, rely on simulated queries and lack disease-specific depth. Moreover, high rubric-based scores do not ensure factual correctness, underscoring the need to assess hallucinations. We developed a human-in-the-loop pipeline to create expert rubrics for de-identified patient questions from the Pancreatic Cancer Action Network (PanCAN). The resulting benchmark, PanCanBench, includes 3,130 question-specific criteria across 282 authentic patient questions. We evaluated 22 proprietary and open-source LLMs using an LLM-as-a-judge framework, measuring clinical completeness, factual accuracy, and web-search integration. Models showed substantial variation in rubric-based completeness, with scores ranging from 46.5% to 82.3%. Factual errors were common, with hallucination rates (the percentages of responses containing at least one factual error) ranging from 6.0% for Gemini-2.5 Pro and GPT-4o to 53.8% for Llama-3.1-8B. Importantly, newer reasoning-optimized models did not consistently improve factuality: although o3 achieved the highest rubric score, it produced inaccuracies more frequently than other GPT-family models. Web-search integration did not inherently guarantee better responses. The average score changed from 66.8% to 63.9% for Gemini-2.5 Pro and from 73.8% to 72.8% for GPT-5 when web search was enabled. Synthetic AI-generated rubrics inflated absolute scores by 17.9 points on average while generally maintaining similar relative ranking.

cs.CL↗
arXiv2026

Spurious Rewards: Rethinking Training Signals in RLVR

We show that reinforcement learning with verifiable rewards (RLVR) can elicit strong mathematical reasoning in certain language models even with spurious rewards that have little, no, or even negative correlation with the correct answer. For example, RLVR training with GRPO improves MATH-500 performance for Qwen2.5-Math-7B by 21.4 percentage points using randomly assigned rewards, nearly matching the 29.1-point gain from ground-truth rewards. To explain this counterintuitive observation, we show that GRPO exhibits a clipping bias from the clip term, which can amplify high-prior behaviors learned during pretraining even without informative rewards. As a case study, we identify one such behavior in Qwen2.5-Math models, which we call code reasoning -- reasoning in code without actual code execution; code-reasoning frequency increases from 65 percent to over 90 percent with spurious rewards. However, the presence of such amplifiable behaviors is highly model-dependent. In practice, spurious rewards that are effective for Qwen models often fail to produce gains for other model families, such as Llama3 or OLMo2. Our results highlight the importance of validating RL methods across diverse models rather than relying on a single de facto choice: large gains can arise on Qwen models even from random rewards that do not reflect genuine capability improvements.

cs.AI↗
arXiv2026

Buy versus Build an LLM: A Decision Framework for Governments

Large Language Models (LLMs) represent a new frontier of digital infrastructure that can support a wide range of public-sector applications, from general purpose citizen services to specialized and sensitive state functions. When expanding AI access, governments face a set of strategic choices over whether to buy existing services, build domestic capabilities, or adopt hybrid approaches across different domains and use cases. These are critical decisions especially when leading model providers are often foreign corporations, and LLM outputs are increasingly treated as trusted inputs to public decision-making and public discourse. In practice, these decisions are not intended to mandate a single approach across all domains; instead, national AI strategies are typically pluralistic, with sovereign, commercial and open-source models coexisting to serve different purposes. Governments may rely on commercial models for non-sensitive or commodity tasks, while pursuing greater control for critical, high-risk or strategically important applications. This paper provides a strategic framework for making this decision by evaluating these options across dimensions including sovereignty, safety, cost, resource capability, cultural fit, and sustainability. Importantly, "building" does not imply that governments must act alone: domestic capabilities may be developed through public research institutions, universities, state-owned enterprises, joint ventures, or broader national ecosystems. By detailing the technical requirements and practical challenges of each pathway, this work aims to serve as a reference for policy-makers to determine whether a buy or build approach best aligns with their specific national needs and societal goals.

cs.CY↗
arXiv2026

Cold-Start Personalization via Training-Free Priors from Structured World Models

Cold-start personalization requires inferring user preferences through interaction when no user-specific historical data is available. The core challenge is a routing problem: each task admits dozens of preference dimensions, yet individual users care about only a few, and which ones matter depends on who is asking. With a limited question budget, asking without structure will miss the dimensions that matter. Reinforcement learning is the natural formulation, but in multi-turn settings its terminal reward fails to exploit the factored, per-criterion structure of preference data, and in practice learned policies collapse to static question sequences that ignore user responses. We propose decomposing cold-start elicitation into offline structure learning and online Bayesian inference. Pep (Preference Elicitation with Priors) learns a structured world model of preference correlations offline from complete profiles, then performs training-free Bayesian inference online to select informative questions and predict complete preference profiles, including dimensions never asked about. The framework is modular across downstream solvers and requires only simple belief models. Across medical, mathematical, social, and commonsense reasoning, Pep achieves 80.8% alignment between generated responses and users' stated preferences versus 68.5% for RL, with 3-5x fewer interactions. When two users give different answers to the same question, Pep changes its follow-up 39-62% of the time versus 0-28% for RL. It does so with ~10K parameters versus 8B for RL, showing that the bottleneck in cold-start elicitation is the capability to exploit the factored structure of preference data.

cs.CL↗
arXiv2026

ALMo: Interactive Aim-Limit-Defined, Multi-Objective System for Personalized High-Dose-Rate Brachytherapy Treatment Planning and Visualization for Cervical Cancer

In complex clinical decision-making, clinicians must often track a variety of competing metrics defined by aim (ideal) and limit (strict) thresholds. Sifting through these high-dimensional tradeoffs to infer the optimal patient-specific strategy is cognitively demanding and historically prone to variability. In this paper, we address this challenge within the context of High-Dose-Rate (HDR) brachytherapy for cervical cancer, where planning requires strictly managing radiation hot spots while balancing tumor coverage against organ sparing. We present ALMo (Aim-Limit-defined Multi-Objective system), an interactive decision support system designed to infer and operationalize clinician intent. ALMo employs a novel optimization framework that minimizes manual input through automated parameter setup and enables flexible control over toxicity risks. Crucially, the system allows clinicians to navigate the Pareto surface of dosimetric tradeoffs by directly manipulating intuitive aim and limit values. In a retrospective evaluation of 25 clinical cases, ALMo generated treatment plans that consistently met or exceeded manual planning quality, with 65% of cases demonstrating dosimetric improvements. Furthermore, the system significantly enhanced efficiency, reducing average planning time to approximately 17 minutes, compared to the conventional 30-60 minutes. While validated in brachytherapy, ALMo demonstrates a generalized framework for streamlining interaction in multi-criteria clinical decision-making.

cs.LG↗