natural language processing engineers

Abstract: Masked Diffusion Models (MDMs) have emerged as a compelling paradigm for language modeling, offering the capability for efficient parallel text generation. However, existing parallel sampling methods typically update multiple masked tokens independently and ignore the complex mutual dependencies among the masked tokens. This independent updating mechanism lacks global coordination and might lead to incoherent outputs. To address this limitation, we propose Representation-based Masked Diffusion Model (RMDM), a framework that leverages the text representation to explicitly encode global semantics and help to parallel update tokens more precisely. Specifically, we first encode text into a continuous semantic space using a pretrained encoder and learn an invertible transformation that normalizes the representation distribution to a Gaussian prior, facilitating efficient sampling during generation. Conditioned on this latent semantic representation, we train a masked diffusion model to learn the conditional text distribution, where the representation serves as global semantic guidance to coordinate parallel token updates and faithfully approximate the target distribution. Empirical results demonstrate that RMDM significantly improves generation quality, particularly in aggressive few-step sampling regimes.

Abstract: A central question in LLM reasoning is whether reinforcement learning (RL) instills genuinely new capabilities or merely reshapes how existing knowledge is expressed during inference. Building on the distribution-sharpening hypothesis, which holds that RL reallocates probability mass toward high-reward trajectories already latent in base models, we ask: can we unlock those latent paths without costly RL fine-tuning? We present Decision-Flow Sampling (DF-Sample), a training-free, data-free inference-time framework that constructs a hierarchical reasoning tree, scores terminal nodes for quality, and back-propagates utilities to inform each intermediate branching decision. Unlike conventional sampling strategies that make purely local step-wise choices, DF-Sample performs explicit global trajectory evaluation before committing to a path, recovering high-quality but low-probability reasoning chains that standard decoding overlooks. On GPQA, DF-Sample achieves 45.6% accuracy, surpassing power sampling (38.9%) and GRPO (39.9%), showing that a training-free method can outperform a trained one. Across three models and four benchmarks, DF-Sample consistently outperforms baselines, indicating substantial latent reasoning potential in pretrained base models.

Abstract: Large language models generate fluent text that can contain unfaithful claims -- a phenomenon known as hallucination. We present a multi-signal detection pipeline combining fine-tuned DeBERTa-v3 classification, Monte Carlo (MC) Dropout uncertainty quantification, and temperature-scaled calibration for response-level hallucination detection. Evaluated on the HaluEval benchmark, our pipeline achieves F1=0.915 and AUROC=0.977 on general-domain tasks, with per-task F1 scores of 0.97 (QA), 0.96 (Summarization), and 0.82 (Dialogue). MC Dropout inference further improves accuracy to 93.2%. A context ablation study confirms the model performs genuine entailment reasoning rather than exploiting surface patterns, with summarization F1 dropping 24% when knowledge context is removed. Learning curve analysis reveals that 25% of training data captures 77% of full-data performance. Beyond detection, we apply Direct Preference Optimization (DPO) to a Qwen2.5-0.5B generator, reducing its hallucination rate from 85.5% to 37.7% (55.9% relative reduction) as measured by our detector. Cross-domain evaluation on the SciFact biomedical benchmark shows that general-domain training transfers poorly (F1=0.52), motivating domain-specific fine-tuning. PubMedBERT fine-tuned on SciFact achieves F1=0.63 and AUROC=0.81, demonstrating that domain-matched pre-training is the strongest adaptation strategy. Code and models are available at https://github.com/varunteja99/hallucination-detection-nlp

Abstract: Per-token gating of forward/reverse KL losses has become a standard technique for on-policy knowledge distillation (OPD), but existing methods such as EOPD (Jin et al., 2026) and ToDi (Jung et al., 2025) each fix a single gating signal and a single gating direction, and the two have never been compared directly. We introduce a four-coefficient parameterization lambda_t = sigma(a * h_t + b * u(x) + c + d * gap_t) in which direction-aligned proxies of EOPD and ToDi appear as one-dimensional (1D) restrictions, and which adds multi-channel composition and an explicit bias as further degrees of freedom. On TweetEval (Barbieri et al., 2020) emotion and hate, with a Qwen3-32B teacher and a Qwen3-4B student, configurations in the full family reach higher accuracy than the matched-magnitude single-channel (entropy-only / gap-only) 1D restrictions in 33 of 36 comparable cells, and a 26-cell mean-match isolation experiment places dynamic gating ahead of effective-KL-matched static baselines in 19 of 26 cells. Because cells share training data, models, and parameter substructure, we report both counts as exploratory aggregate directional evidence rather than as independent hypothesis tests. Targeted three-seed paired replications of the nine headline comparisons singled out by that sweep -- including a third task, offensive -- are directionally consistent, but individually smaller than the single-seed estimates and not significant at n=3. We therefore present the parameterization primarily as a shared coordinate system for comparing per-token gating designs in short-output classification OPD.

Abstract: Test-time adaptation has emerged as a lightweight alternative to costly post-training for improving the reasoning capabilities of Large Language Models (LLMs) on downstream tasks. Predictive entropy provides a model-derived signal for such adaptation, guiding models toward higher-confidence reasoning states without external verifiers or reward models. However, higher confidence does not necessarily imply correctness, as LLMs may remain highly confident along incorrect reasoning trajectories. We observe that high-confidence reasoning is more likely to be correct when confidence remains stable under local perturbations. Based on this observation, we propose Test-Time Adaptation via Stability-Aware Confidence Optimization (TASCO), a framework that incorporates local stability into confidence-based test-time adaptation while keeping the LLM frozen. TASCO operationalizes local stability by optimizing a lightweight task-level prefix under two alternative perturbation strategies: Random Perturbation promotes distributional stability across trajectories induced by nearby perturbed prefixes, whereas Sharpness-Aware Perturbation targets worst-case local sensitivity. Experiments demonstrate that TASCO improves reasoning accuracy and token efficiency across diverse LLMs and reasoning benchmarks, while behavioral analyses show that it maintains stable confidence under local perturbations without prematurely concentrating the model's predictive distribution.

Abstract: A wrong answer does not show whether the model lacked the needed information or held it and failed to use it. On an entity-obligation binding task, a language model can emit an incorrect prompt-supplied binding while a linear probe can recover the correct one from its frozen hidden state. We measure how often this occurs across 16 public checkpoints, each evaluated with three seeds. We fit a probe on a training fold, select its layer on a validation fold, and report results on a disjoint test fold. On the trials each model gets wrong, probe accuracy exceeds the strict present-obligation baseline, 1/K = 0.125, by +0.196 (95% CI [+0.101, +0.296], bootstrapped over models). A query-entity counterfactual rules out token presence and recency. A score built from the sign of probe-output disagreement improves failure detection over the model's own confidence by +0.079 AUROC (95% CI [+0.036, +0.126]). Raw probe confidence gives no measurable improvement over model confidence. Steering the residual stream toward the probe-decoded binding, with no gold label, raises accuracy on all eight models tested by a mean of +0.168 (95% CI [+0.066, +0.280]). Where recent studies report that probe-detected errors are resistant to interventions, we find that in-context binding is a setting in which probes are actionable.

Abstract: Large language models are increasingly used as judges to measure social bias in text, yet the passages they judge are often noisy, containing typos, informal spelling, and broken punctuation. The consequences of such surface noise for social bias measurement remain unclear. To investigate this question, we apply five realistic noise conditions at multiple intensity levels to 3,822 stereotype-related responses and compare the resulting bias judgments with those on the original text. We find that such surface noise does not degrade bias measurement symmetrically: it is far more likely to turn neutral judgments into biased ones than biased judgments into neutral ones, by up to a 120x margin. We further observe two non-obvious effects across four LLM judges: in the most fragile judge the distortion is at its purest at mild, realistic noise levels, where erasure is scarcest, and as judges grow robust it attenuates toward parity rather than reversing. Bias measured on noisy text is therefore systematically overestimated, most in the categories that matter most for fairness.

Abstract: Large language models are typically trained under uniform token weighting, which allows frequent and low-information tokens to dominate learning and can increase the tendency to memorize surface-level text spans. To address this, we present an information-weighted cross-entropy loss that rescales token-level contributions using TF-IDF statistics, emphasizing semantically informative tokens while down-weighting ubiquitous ones. Experiments on five decoder-only LLMs ranging from 1.1B to 13B parameters show consistent reductions in memorized substring length while preserving perplexity and downstream task performance. Under LoRA fine-tuning, TF-IDF reduces average substring memorization length by 14% across all five models. Under full-weight fine-tuning on TinyLLaMA 1.1B, the reduction reaches 58%. Our approach is architecture-agnostic and can be incorporated into existing training pipelines with less than 3% computational overhead, offering a lightweight and principled way to mitigate memorization without disrupting standard training dynamics.

Abstract: Verbalized confidence, long dismissed as overconfident, coarse, and prone to round-number clustering, is now the more robust soft-scoring mechanism for LLM-as-a-Judge on top-tier proprietary models. Across SummEval, AggreFact, and HelpSteer2, spanning up to 18 LLMs, we show that the standard advice to prefer log-probabilities no longer holds on post-2025 models, where verbalized confidence is the better signal. We call this a compatibility shift. On top of a standard verbalized-confidence baseline, we introduce two new ingredients: an overconfidence advisory and self-debate. Together they improve calibration, score-distribution spread, and robustness to task subjectivity. We further observe a generation effect: post-2025 models accommodate these two additions with little balanced-accuracy cost, whereas pre-2025 models pay a measurable penalty. Compared with logprob-based G-Eval, verbalized confidence is the more subjectivity-robust soft signal on GPT-family top-tier releases. The shift is invisible under accuracy-only reporting. Rather than defaulting to hard predictions, we recommend broader use of soft scoring in LLM-as-a-Judge. More broadly, verbalized confidence has moved from a weaker substitute for logprobs to a practical soft-scoring mechanism for contemporary LLM judges.

Abstract: Learning from limited text requires models to use context, generalize to new inputs, and retain useful capabilities. Qiushi Engine conducted a long-horizon, end-to-end autonomous research program on BabyLM 2026 Strict-Small, within 10 million corpus words and 100 million cumulative word presentations. Three stages connected frontier advancement, principle discovery, and principle-guided model improvement. Stage I combined compact restatements, budget reinvestment, and residual incremental learning to build a frontier model. Stage II found that exact repetition and aligned restatement produce different patterns of context use, depending on target relations and prediction windows. In controlled tasks, recovering familiar performance did not ensure that unseen inputs could still use learned computations. These findings support a testable data-efficient learning principle: organize experience around the contextual dependencies needed for prediction; separately design visible information, supervision, and preservation; test learning, generalization, and retention. Stage III retained source text, masked more local clues, supervised selected targets, and preserved predictions on ordinarily masked inputs. Two continuation seeds from the same parent outperformed ordinary continuation on the complete nine-metric aggregate. Overall rose from 42.02 to 42.25 across two generations; the second achieved the highest Overall in the public Strict-Small snapshot of 8 September 2026. Further studies addressed compression, relational anchors, shared representations, and measurement. Models are available on Hugging Face; code and research records accompany the GitHub repository. Together, these stages illustrate Research RSI: recursive self-improvement of the research process. Scientific understanding and method innovations change subsequent questions and designs; new experiments test and refine them.

Abstract: Language models under one million parameters matter for edge deployment, domain adaptation, and reproducible research, yet a two-layer LSTM or Transformer at embedding width d = 128 still spends roughly one third of its capacity on the output matrix W_out in R^(d x |V|). We propose Riemannian Language Models (RiLM), which remove that layer entirely: context unfolds as a trajectory on a Riemannian manifold, and next-token probabilities arise from squared geodesic distance between the current state and vocabulary embeddings. The same embedding map serves input and output -- decoding is geometry. We instantiate the framework on flat R^d (Flat RiLM) and the Poincare ball H^d (HypRiLM) with a shared MLP composition map phi (~290k parameters, d = 128, |V| = 2000). Across five seeds on WikiText-2, HypRiLM reaches 54.2 +/- 0.2 validation perplexity versus 87.6 +/- 0.6 for Flat RiLM; tied and matched LSTM, Transformer, and SSM controls remain at 113-147 PPL on WT-2 -- HypRiLM leads by roughly 2x over the strongest tied recurrent baseline (SSM, 113.0 +/- 3.8). Penn Treebank and a 10k-vocabulary stress test confirm that geodesic decoding transfers across corpora and larger |V|, while hyperbolic curvature helps selectively. We also characterize boundary collapse in naive hyperbolic recurrence and show how Mobius stabilization restores trainability. Claims are scoped to controlled small-model comparisons, not full-vocabulary state of the art.

Abstract: Arabic morphology remains challenging for large language models, since fluent generation does not guarantee accurate morphosyntactic control. Existing Arabic evaluations mainly target downstream tasks and do not directly test controlled morphological generation from explicit lexical and feature-based input. We introduce YallaMorph, a large-scale benchmark for Arabic morphological generation covering verbs, nouns, adjectives, their cliticized forms, and invalid configurations. We evaluate multilingual and Arabic-oriented LLMs under diacritized and undiacritized settings over 600K benchmark entries. Results show that Arabic morphological generation remains difficult, especially for cliticized, unseen, and morphologically rare forms.

Abstract: When training data are limited, increasing parameter count is not the only way to improve language-model performance. A small parameter set, when repeatedly applied, can also deliver comparable performance. We study Looped GPT-BERT in the BabyLM 2026 Strict-small setting, combining GPT-BERT's masked next-token and causal language-modeling objectives with depth-wise parameter sharing. We train on a preprocessed 7.48M-word English corpus and compare objective ratios, non-looped and looped architectures, and loop counts. Our final $4\times12$ model uses four physical layers for twelve recurrent traversals and contains 12.18M parameters. The BabyLM 2026 leaderboard reports an Overall Average of 35.42 and an NLP Average of 48.48. Compared with public BabyLM 10M Strict-small GPT-2 and GPT-BERT baselines, it achieves comparable performance on selected linguistic and downstream metrics, including BLiMP and GLUE, with fewer parameters. The loop ablations show that additional recurrent computation can improve training and preserve strong performance on selected linguistic tasks, whereas poorer performance on other tasks may reveal an inherent limitation of the looped design: using only a few physical layers restricts the model's representational space.

Abstract: This study examines the compositionality of steering vectors for language and behavioral control in large language models. Focusing on language, jailbreak, and conciseness, we investigate whether additive, training-free composition of attribute steering vectors can preserve the intended steering effect of each attribute, across four instruction-tuned models from two model families and two size scales. We find that single-attribute steering is reliable for all three attributes, but only within an appropriate combination of intervention layer and steering strength, with abstract behaviors (jailbreak, conciseness) favoring middle layers and language favoring earlier layers. We show that additive composition of two attribute vectors succeeds in steering both attributes simultaneously when each is injected at its own best-performing layer, and that this partially extends to three simultaneously composed attributes, addressing an inconsistency left open by prior work on training-free composition. We further analyze the geometric properties of these steering vectors, finding that they are approximately orthogonal in the residual stream, consistent with their compositional behavior.

Abstract: A dense retriever encodes a question as one vector, but the question arrives one word at a time. We read 2,144 held-out headlines from Thu Vien Phap Luat (Vietnamese legal library) word by word with Nemotron-3-Embed 8B/1B and Qwen3-Embedding 8B/0.6B, encoding 65,444 prefixes against 20,034 articles, plus every prefix of 3,438 sub-questions from 1,112 multi-question headlines and of 168 answers. (i) The gold article becomes rank 1 after a median of 6-7 content words in every encoder, before the interrogative frame is read, and stays there to the end in 78-85% of cases. (ii) In a multi-question headline the lock is inside the first sub-question 94-98% of the time; the second leaves rank unchanged in 89-95%; encoded alone, the second reaches rank 1 in 42-58% vs 91-96% for the first, at the same lock word (95-97% identical). (iii) Numbers, dates and instrument identifiers move the embedding twice as far as content words and four times as far as interrogative words; 72-78% of steps move toward the gold article, and the closing interrogative frame moves against that direction in 95-99% of headlines. (iv) Rank/cosine clustering yields six archetypes (instant, typical, unstable, late, never-locking) that differ by legal area and form (chi-squared p < 1e-8): real-estate and litigation headlines never lock on a number; environmental and accounting headlines do so a third of the time. (v) An answer read word by word retrieves its article after 8-16 words and addresses the sub-questions in order asked in 83-89% of cases. (vi) A word's step keeps a consistent direction across headlines (cosine 0.25-0.33; 0.44-0.60 for numbers); a preceding question rotates that step by about 60 degrees and a greeting by about 30 degrees; steps shrink as i^{-0.8}; and a two-question headline is within 12-17 degrees of a linear mix of its two questions. We call this a context-modulated additive walk.

Abstract: Automatic translation quality metrics trained on general-domain corpora systematically fail on social media content, where communicative intent is encoded in culturally loaded expressions (internet slang, homophonic ciphers, and platform-specific idioms) rather than surface token patterns. We conduct a systematic empirical analysis demonstrating that standard metrics including COMET, XCOMET, and BERTScore exhibit near-zero or negative correlation with human cultural judgments, and even display a severity inversion in which scores increase as translation quality deteriorates. We further show that this failure extends to large language model judges: Qwen3-235B achieves Cohen's kappa of only 0.162, revealing that the bottleneck is not reasoning capacity but cultural grounding: models lack the domain-specific cultural knowledge needed to identify which aspects of a translation require scrutiny. To address this, we propose CuRIL, a reinforcement learning framework that internalizes cultural reasoning: cultural annotations are prepended inside the model's reasoning, excluded from policy gradients via a token-level loss mask, and injected with a probability that decays to zero over training, progressively forcing autonomous cultural judgment. On a 1,444-sample human-annotated social media translation benchmark, Qwen3-8B trained with CuRIL achieves Cohen's kappa 0.370 and Exact Match accuracy of 45.22%, approaching Gemini-3.1-Pro with 30x fewer parameters and surpassing models up to 235B in scale. We further demonstrate that our judge produces reliable reward signals for downstream translation optimization, reducing the low-quality translation rate by over 20 percentage points under independent human evaluation.

Abstract: This paper presents a small-scale quantitative experiment that links syntactic structure to stylistic functions in narrative prose. Starting from a dependency-parsed corpus of 3,300 sentences, we derive sentence-level stylistic labels across five categories --- descriptive, introspective, causal, ideological, and neutral --- using a transparent rule-based procedure that inspects lemmas, universal part-of-speech tags, and syntactic relations. For each sentence we construct a compact representation of its syntactic profile as a sequence of linearised triples combining lemma, POS tag, and dependency relation. These patterns serve as input to standard machine-learning classifiers trained to predict sentence-level style. The best-performing model achieves a macro-F1 of 0.948 under 10-fold cross-validation. The experiment is implemented entirely in Python using open-source tools. Our goal is not to propose a fully fledged stylistic theory, but to offer a reproducible and extensible workflow for exploring how grammatical structure contributes to narrative interpretation.

Abstract: Rotary position embedding (RoPE) uses each token's integer position to determine the rotation applied inside attention. This works well for local token order, but increasing context length creates a positional train-test mismatch: RoPE produces relative rotations at offsets not seen during training. Methods that rescale, interpolate, randomize, or bias positions specify how attention handles those offsets, but still derive positional information from a growing token counter. We instead divide a token stream into units, retain ordinary RoPE positions within each unit, and assign every completed unit an address computed from its content. Adding units then applies the same learned map to new content rather than extending a positional range or an identifier table. We prove that this construction preserves local RoPE exactly, leaves the attention comparison between two fixed tokens unchanged when other units are inserted or reordered, and does not create new relative rotations merely because more units are added. In a character-level Tiny Shakespeare diagnostic, a model trained on 256-character contexts has validation perplexity 4.04 at 256 characters and 3.82 at 4096, while continuous RoPE changes from 4.71 to 12.09. A second diagnostic shows that content-based addressing can retrieve and use information from multiple serialized facts. These are controlled shallow experiments, not scale benchmarks, but they support a direct prescription: use position to address locally and content to address across units.

Abstract: Diffusion Large Language Models (dLLMs) generate text via bidirectional iterative denoising, naturally supporting user-specified constraints anchored at arbitrary output positions, a paradigm known as In-place Prompting (IPP). We formalize this as the In-place Instruction Following (IIF) task and construct IIF-Bench, a hierarchical benchmark spanning literal, style, and discourse-function constraints, paired with a rubric-based local-global evaluation protocol. An inference-time attention-bias probe suggests that vanilla dLLMs often under-prioritize constraint spans during denoising. We then propose GRAFT, an IPP-oriented post-training framework combining constraint-aware SFT and preference optimization. On four representative dLLMs, GRAFT raises the average IIF score from 57.75 to 73.10 (+15.35 points), with absolute gains of 15.91 and 15.57 points on literal and discourse-function constraints, while preserving general generation ability.

Abstract: This paper presents our system for CCL2026-Eval Task 5: Minor-Grain Breeding Information Extraction (MGBIE), which jointly extracts 12 entity types and 6 relation types from minor-grain breeding literature. We propose RAME (Retrieval-Augmented Multi-Prompt Ensemble), a training-free framework that elicits multiple LLM outputs under controlled diversity and aggregates them by majority voting to obtain high-confidence predictions. RAME combines (i) retrieval-augmented few-shot selection via a hybrid BM25-embedding retriever, (ii) a three-prompt ensemble (Strict, Relaxed, Balanced) spanning the precision to recall spectrum, and (iii) large-scale repeated sampling with majority voting to filter noisy predictions. Built on DeepSeek-V4-Flash, RAME achieves a Total Score of 0.499 (NER 0.730, RE 0.346) on the leaderboard, ranking 1st and surpassing the official Track-A baseline powered by GPT-5.5 (0.448), representing an 11.4% relative improvement. Code is available at https://github.com/king-wang123/CCL26-RAME.

Abstract: Activation steering has emerged as a lightweight, inference-time approach to control the behavior of Large Language Models (LLMs). However, traditional steering vectors used to intervene in LLMs' activations, such as those derived from the difference-in-means method, tend to entangle multiple semantic and stylistic concepts into a single composite direction, leading to unpredictable steering effects. Our core objective is to disentangle this composite direction into its constituent concepts. To this end, we propose Steering Vector Dissection, a framework to explicitly isolate individual and semantically consistent features from these composite directions. Specifically, we pair positive and negative activations and take their differences to generate a set of instance-level steering vectors, and train a dedicated Sparse Autoencoder (SAE) directly on them. Quantitative evaluations across two datasets, two models, and two intervention depths show that our method yields a set of semantically consistent basis vectors whose steering effects are mutually distinguishable. Furthermore, we show that this disentanglement enables precise control over model behaviors.

Abstract: Cantonese is widely spoken but remains low-resource in written data, with no large corpus of native Cantonese reasoning traces available for model training. We develop and release CantoneseLLM v2, comprising models based on Qwen3 8B and 30B-A3B. The models are trained through CPT on 784 million Cantonese and Hong Kong-related tokens, chat-vector merging, SFT, DPO, and RLVR. Evaluation across the training stages shows that chat-vector merging transfers instruction following but preserves the donor model's reasoning language, while SFT with limited Cantonese reasoning data substantially shortens or removes reasoning traces and reduces benchmark performance. DPO restores the reasoning-block format, particularly for the 8B model, but recovers only part of the lost performance. The RLVR training with Cantonese language and Traditional Chinese scripts as multiplicative constraints introduced Cantonese language alignment and restored the lost performance. The 30B-A3B model reaches 73.16 on HKCanto-Eval, within 1.20 points of its merged checkpoint, while retaining the Cantonese reasoning behaviour absent from that checkpoint. We release the model checkpoints, the training environments, and a thirteen-year Traditional Chinese Common Crawl dataset. The models can be accessed at https://huggingface.co/collections/hon9kon9ize/cantonesellm-v20