Search: “paraphrase-invariance”

Smart search ranks by meaning as well as keywords (one row per story, last 45 days).

Same Trajectory, Contradictory Rewards (ROBORMBENCH): Paraphrase Fragility in Vision Language Reward Models

New ROBORMBENCH benchmark shows vision-language reward models can flip robot success/failure judgments when goal instructions are paraphrased.

The authors show that paraphrasing the instruction alone can substantially change progress scores from VLM reward models, even flipping identical robot trajectories between failure and success. ROBORMBENCH comprises 2,390 real-robot trajectories with ground-truth progress labels and 21,673 verified paraphrases covering lexical, syntactic, and action-goal rewrites. Instability is widespread across proprietary and open-source VLMs, grows with more divergent rewrites, and is not reliably reduced by scale or explicit reasoning, while trajectory-grounded dedicated reward models are markedly more stable.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 17:47 UTC · 12d agoAI research1

35

I Am No One: Style-Aware Paraphrasing for Text Anonymization

Prompt-driven style-aware paraphrasing with LLMs cuts authorship attribution F1 by 60-70% while preserving content quality.

The paper proposes a style-aware, prompt-driven anonymization approach using pretrained LLMs to build compact stylistic profiles from minimal samples and rewrite text to suppress identifiable style markers while preserving meaning. It addresses stylometric re-identification risks in anonymized text, including ASR transcripts of meetings and call-center calls where leakage persists after acoustic anonymization. Across blog and review datasets, the approach reduces authorship attribution F1 by 60-70%, substantially outperforming both DP-based and non-DP baselines while maintaining readability.

arXiv cs.CR·Sep 11, 02:07 UTC · 5d agoResearch

40

Verifiable by Construction: Claim-Level Evaluation of Verbatim Citation in Clinical Question Answering

Evaluation of twelve LLMs on 222 clinical questions shows verbatim quotes rarely substantiate claims; claude-opus-5 fully substantiates only 37.1%.

The authors build a standardized harness over four clinical practice guidelines and evaluate twelve LLMs on 222 synthetic clinical questions, measuring citation attachment, verbatim quote production, and claim substantiation. Most models attach verbatim quotes to over 90% of claims from prompting alone, though lightweight models like claude-haiku-4.5 struggle. Quotes frequently fail to substantiate claims: claude-opus-5 quotes 98.0% of claims but fully substantiates only 37.1%, exposing a capability gap for verifiable clinical QA.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 17:53 UTC · 2d agoAI research

38

Inoculation Midtraining with Learned Neologisms

Inoculation Midtraining confines unsafe LLM behavior to a neologism-marked context, reducing misalignment after unsafe post-training but leaking under nearby contextual cues.

The paper introduces Inoculation Midtraining, which teaches a base model during midtraining that unsafe behavior belongs to a context marked by a learned neologism token, then post-trains on unsafe data within that context. Across supervised fine-tuning and RL post-training regimes, the technique reduces misalignment while preserving transfer of benign properties like German or Shakespearean prose. However, it does not outperform standard Inoculation Prompting, is sensitive to training configuration, and produces a leaky boundary that nearby contextual cues can reactivate. The authors conclude it is not yet a load-bearing component of a developer safety framework.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 17:12 UTC · 2d agoAI safety & security

42

LexFlip: A Dissociation Diagnostic for Legal Meaning Preservation Metrics

LexFlip releases 373 minimal perturbations of Quebec statutory French that reverse legal force while preserving tokens, exposing weaknesses in embedding-based meaning preservation metrics.

LexFlip provides 373 minimal perturbations of Quebec statutory French that reverse legal force while preserving 0.93 of tokens, creating dissociation items that break monotone token-overlap metric validation. The seven embedding and BERTScore metrics tested register only 0.022-0.039 of their identical-to-unrelated range on these edits, versus 0.670 for bidirectional NLI. Against FrJudge, with a measured human ceiling of r=0.597, a bare length feature outscores every semantic metric tested.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 15:49 UTC · 12d agoAI research

22

Technical Manual for a Toolkit for Measuring Contextual Individuation in Transformer Language Models

An open methodology toolkit measures whether transformer language models contextualize fixed word forms across domains using bridge forms and layer-wise silhouette analysis.

The manual documents an open toolkit built around 'bridge forms' - identical written words recurring across two or more subject domains with a different sense in each - to test whether transformer language models individuate word occurrences by context beyond the embedding layer. It covers declarative specification of bridge forms, Wikipedia corpus acquisition, occurrence localization, layer-wise representation extraction, domain-pairwise silhouette measurement, and visualization, justifying each choice against failure modes such as sense contamination and subword-tokenization misalignment. It is a methodological and implementation reference and reports no empirical results.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 16:38 UTC · 12d agoAI research

14

Convergent Emergence of In-Context Learning Across Modalities

Controlled experiments show few-shot in-context learning emerges across six modalities including language, genomes, images, and proteins, partially supporting a convergence hypothesis.

The paper tests the Convergent Emergence Hypothesis: that few-shot in-context learning, when it emerges, shares a common cross-modality difficulty profile. A controlled framework instantiated the same task suite across six modalities: language, genome, integer sequences, time series, images, and proteins. Paired-mapping ICL emerged in all six modalities, surpassed controlled baselines, and showed correlated per-task effects in five of them, providing partial support for the hypothesis.

Hugging Face daily papers·Sep 11, 20:00 UTC · 5d agoAI research

30

Unifying Conformal Language Tasks with In-Context Ensembles

Researchers propose Conformal Relevance, which builds conformal score functions via in-context example curation and ensembling to improve conciseness across seven NLP tasks.

The paper targets NLP tasks like summarization and extractive QA that reduce to retrieving content under coverage and conciseness constraints. Conformal Relevance replaces hand-engineered LLM scoring prompts with curated in-context examples and ensembles, maintaining coverage guarantees while improving conciseness with minimal manual input. The authors demonstrate the framework on seven NLP tasks and contribute theory, including a complementarity condition for when ensembling improves worst-case sentence scores and a saturation bound on ensemble gains.

Hugging Face daily papers·Sep 1, 20:00 UTC · 15d agoAI research1

42

RelateAnything: Real-Time Open-Vocabulary Relation Prediction From Any Inputs

RelateAnything is a 53M-parameter open-vocabulary relation prediction model running at 20 ms/frame, with 2.3-3.5x higher mean recall than comparable open-vocabulary methods.

RelateAnything predicts scored relations between image regions using any predicate vocabulary supplied at inference as text embeddings, with object labels never required as input, so region sources can change without retraining. Training covers 19,103 predicates using positive-unlabeled supervision; the authors release RA-4M (474k images, 4.3M geometrically verified relations over 10,102 free-text predicates) and the OV-SGG-Bench evaluation suite. The 53M-parameter model runs at 20 ms/frame and achieves 2.3-3.5x the mean recall of the strongest comparable open-vocabulary method across cross-dataset and zero-shot benchmarks. Model, corpus, and benchmark are public.

Hugging Face daily papers·Sep 10, 20:00 UTC · 6d agoAI research1

40

Molecular Déjà Vu: Digit-Level Retrieval of Published Values in Frontier Language Models

Audit of 22 frontier models finds widespread verbatim retrieval of published molecular property values, with higher reasoning increasing recall of memorized numbers.

An arXiv audit tests 22 frontier LLMs across 12 molecular regression benchmarks for verbatim retrieval of published values. More than 50% of the LLMs show verbatim retrieval on five datasets, and identical experiments are flagged 89% more often at a high reasoning level than at the lowest one. Suppressing retrieval moves model prediction errors closer together in relative terms, suggesting predictive capability is not determined solely by memorized values.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 17:32 UTC · 12d agoAI research1

42

Structural Jailbreaks Generalize but Do Not Compound: A cross-provider and multilingual study of Involuntary In-Context Learning

Researchers show IICL structural jailbreaks generalize to Google Gemini, lifting attack success to 80-100% on harm and financial benchmarks; non-English prompts attenuate it.

The study red-teams two Google Gemini models with Involuntary In-Context Learning (IICL), a structural jailbreak reframing harmful requests as the final cell of a data-labeling task. IICL lifts attack success from at most 6.7% to 80-90% on HarmBench and 97-100% on financial abuse (FinProof), an order of magnitude above prior results on OpenAI's GPT-5.4. Against a compounding hypothesis, forcing IICL output into Spanish, Hindi, or Arabic attenuates the attack in 11 of 12 conditions, attributed to a 'relevance curse' producing lower-quality harmful content in lower-resource languages. Findings replicate under an independent non-Google judge (Cohen's kappa 0.86 over 377 paired verdicts).

arXiv cs.CR·Sep 8, 07:38 UTC · 8d agoAI safety & security

58

Studying Image Tokenizers as Visual Languages in Unified Multimodal Models

A controlled autoregressive testbed shows validation losses must be analyzed per task, and image tokenizer choice affects joint multimodal text modeling.

Researchers built a pure-autoregressive testbed to study image tokenizers as the 'visual language' of unified multimodal models, tracking task-specific validation losses during multimodal continual pretraining across text, image, text-to-image (T2I), and image-to-text (I2T) prediction. They found that losses exhibit distinct scaling behavior per task and rank tokenizers differently, and that I2T loss over a shared text vocabulary gives a more consistent loss–performance signal than T2I loss. Better reconstruction does not necessarily yield lower task-specific losses or stronger downstream performance, and tokenizer choice can affect text modeling under joint optimization. Case studies examine the discriminator, semantic supervision, and vocabulary size design axes.

Hugging Face daily papers·Sep 7, 20:00 UTC · 9d agoAI research1

40

Large Language Models Develop Belief State Geometry In-Context

Probing six open-source LLMs on HMM-generated data shows belief states are linearly decodable (R² 0.83–0.99), suggesting in-context learning approximates Bayesian prediction.

Researchers prompted six open-source LLMs with data from 40 hidden Markov models selected for non-trivial belief structure and probed residual-stream activations for belief states (posteriors over hidden states). Belief states were linearly decodable with peak R² values of 0.83–0.99 across HMM/LLM combinations, spanning early to late layers. Patching and steering the probe-identified subspace preserved downstream prediction quality while control interventions degraded performance substantially, establishing functional relevance. The results provide representation-level evidence that in-context learning approximates optimal Bayesian prediction over a context-inferred generative model.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 16:17 UTC · 1d agoAI research1

45

Studying Image Tokenizers as Visual Languages in Unified Multimodal Models

A controlled pure-autoregressive testbed shows task-specific validation losses rank image tokenizers differently, with I2T loss the most consistent signal.

Researchers built a controlled pure-autoregressive testbed and tracked task-specific validation losses during multimodal continual pretraining across text, image, text-to-image (T2I), and image-to-text (I2T) prediction. They find losses should be analyzed per task because they exhibit distinct scaling behavior and rank tokenizers differently, and that the loss-performance relationship depends on the predicted token space. I2T loss, computed over a shared text vocabulary, correlates consistently with both generation and visual understanding performance after supervised finetuning. Case studies revisit the discriminator, semantic supervision, and vocabulary size as tokenizer design axes.

arXiv cs.AI / cs.LG / cs.CL·Sep 8, 17:57 UTC · 8d agoAI research1

30

Bag of Tricks or Bag of Myths? Reducing Modeling Complexity with Task Knowledge in Explainable Suicide Risk Assessment

Researchers audit 31 NLP techniques for clinician-annotated suicide risk prediction, finding only 5 of 31 comparisons yield reliable gains.

A study of 1,635 clinician-annotated social media posts ran roughly 300 controlled experiments across 7 methodological families, auditing techniques such as model scaling, synthetic data, ensembling, and threshold tuning under severe class imbalance. The proposed system reformulates risk factor prediction as entailment between posts and codebook definitions, using architecturally diverse ensembles with class-balanced training and deployment-consistent calibration. It scores 0.8203 for risk, 0.7953 for evidence, and 0.7045 macro-F1 for factors, ranking third among 53 teams.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:14 UTC · 9d agoAI research

14

Type Diversity Enables Transformers to Generalise Compositionally

Researchers show lexical-versus-structural compositional generalization gaps in Transformers stem from type diversity imbalance in datasets, not architectural limits.

The paper argues that Transformers' difficulty with structural compositional generalization is an artifact of low structural type diversity in prior benchmark datasets rather than an architectural limitation. Using Grammatical Framework, the authors create linguistically diverse variants of COGS and SLOG. They find type diversity correlates with compositional generalization equally in lexical and structural test cases, contradicting previous claims that compound divergence explains task difficulty.

arXiv cs.AI / cs.LG / cs.CL·Sep 11, 17:59 UTC · 5d agoAI research

20

Does Syntax Matter? A Graph-Augmented Variational Topic Model for Computational Social Sciences

SCPTM graph-augmented variational topic model shows syntax aids topic diversity and descriptor quality but gains stem mainly from the variational encoder.

The Structural Contextual Probabilistic Topic Model represents corpora as heterogeneous document-word graphs with lexical and syntactic edges processed by a Graph Attention Network inside a VAE for mixed-membership topic distributions. Across four corpora, neural gains in document-topic alignment are attributable to the variational encoder rather than syntax, while graph-augmented variants improve topic diversity everywhere. Dependency paths add value on argumentative deliberative texts but are redundant in technical and institutional registers.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:39 UTC · 9d agoAI research1

21

NCP-ArchPreview Technical Report: Moving towards Latent Space Language Models through Next Concept Prediction

An 8.9B-parameter latent-space language model using next-concept prediction matches OLMo-3-7B pretraining loss with only 51.3% of the training tokens.

NCP-ArchPreview augments next-token prediction with Next Concept Prediction over a product-quantized concept vocabulary built from hidden states, trained jointly end-to-end. The 8.9B model was trained on 5.73T tokens from the Dolma-3 dataset, the largest latent-space language model demonstration to date. It consumes 51.3% of the tokens to reach OLMo-3-7B's final pretraining loss and outperforms it by 2.45 points on the downstream macro-average, including a 5.99-point GSM8K gain. The learned latent space also enables lightweight domain adaptation via a 17M-parameter VQ module and improves speculative drafting accepted length by 4.17%.

Hugging Face daily papers·Sep 8, 20:00 UTC · 8d agoAI research1

52

Encoded Early, Used Late: Where Transformers Begin to Act on an Inferred Partner's Expertise

Probing finds transformers represent an inferred dialogue partner's expertise in early layers long before it causally influences output.

Using ExpertCollab, a corpus of multi-turn research-planning dialogues between model-played personas at four expertise levels, researchers show that a partner's inferred expertise is most decodable in early transformer layers and decays to near chance before the network's midpoint. Counterfactual patching reveals that injecting the expertise difference at peak decodability barely changes a fixed late-layer readout, while injection past the midpoint propagates almost completely. The result bounds where readout or steering of partner-conditioned behavior must intervene, demonstrated on a single model with a synthetic corpus.

Hugging Face daily papers·Sep 6, 20:00 UTC · 10d agoAI research

30

Climate-ModernBERT: Revisiting Corpus Composition for Domain-Adaptive Continued Pretraining

Climate-ModernBERT domain-adapted encoders reach 76.3 average F1 across nine climate benchmarks, 2.8 points above vanilla ModernBERT-Base.

The authors continue pretraining ModernBERT-Base on three climate corpora - academic text, climate-filtered web data, and synthetic documents - and compare joint mixtures against parameter-space merging of specialized checkpoints. The best model achieves 76.3 average F1 across nine climate NLP benchmarks, a 2.8-point improvement over the vanilla baseline. Academic climate corpora provide the strongest adaptation signal, and parameter-space merging outperforms joint multi-source training while preserving complementary corpus information; all variants are released.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:41 UTC · 9d agoAI research

22

Right Tool, Right Job: Native-Language Evaluation, Tokenizer Sensitivity, and Methodological Findings from a French-Only BabyLM

French BabyLM entry METRON-FR (125M GPT-2, 92.47M words) shows tokenizer artifacts dominate child-scale zero-shot evaluation; proposes standard diagnostics.

METRON-FR is a 125M-parameter GPT-2 pretrained on 92.47M French words, submitted to the BabyLM 2026 Strict track, scoring 85.97% on the native Quebec-French QFrBLiMP benchmark and 62.80% on the BabyLM-weighted leaderboard. A cross-lingual GLUE protocol combining French task-data translation with rank-16 LoRA shows relational tasks gain while world-knowledge tasks regress. Bilingual Lexicon Induction reaches p@1 of 68.84%, 18x above chance, and ablations show single-token zero-shot scoring is dominated by tokenizer and template artifacts at child scale.

arXiv cs.AI / cs.LG / cs.CL·Sep 15, 16:43 UTC · 1d agoAI research

12

Building Multilingual Bridges: Data Mixing as the Pillar of Generalization for In-Language Reasoning

Tiny Aya L2-Thinker, a 3.35B model, achieves over 93 percent in-language reasoning across 60 languages via optimized multilingual data mixing; weights released.

The paper studies L2 reasoning, the ability to reason consistently in the language of the user's prompt, approached through SFT data composition and scheduling. Tiny Aya L2-Thinker (3.35B) achieves an in-language reasoning rate above 93 percent across 60 languages on six benchmarks spanning math, commonsense, instruction following, open-ended generation, and cultural reasoning. Findings show generalization to held-out languages comes from broader language coverage, multilingual non-reasoning data, and a strong English reasoning backbone, suggesting reasoning is language-agnostic and transferable without per-language supervision. Model weights and multilingual reasoning data are publicly released.

arXiv cs.AI / cs.LG / cs.CL·Sep 9, 16:56 UTC · 7d agoAI research1

45

SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking

SAS trains attention sparsification end-to-end with the language modeling loss, beating sparse attention baselines especially under tight context budgets.

Simple Attention Sparsification (SAS) injects the selector's continuous scores into attention logits in log form inside the softmax, letting gradients from the language modeling loss directly update the ranking of context units. The method uses normalized softmax gates calibrated against the current block and a memory-efficient Triton kernel integrated into FlashAttention-style computation. Across reasoning, long-context, and agentic tasks, SAS consistently outperforms trainable sparse attention baselines across budgets, with the largest gains under tight attention budgets.

Hugging Face daily papers·updated Sep 11, 17:58 UTC · 5d ago·first Sep 10, 20:00 UTC · 6d agoAI research 2 sources1

30

From Reweighting to Rewriting: Unlocking the Intervention Effects of Influential Samples in Training Data Attribution

Study shows rewriting responses of influence-selected training examples shifts LLM behavior more strongly than reweighting the same samples.

The paper examines training data attribution, arguing that influence functions identify high-leverage examples whose value goes unrealized under conventional weight-based reweighting interventions. It introduces influence-guided response rewriting, which replaces the responses of influence-selected examples with behavior-aligned or behavior-opposed supervision while keeping instructions fixed, tested across four open-weight LLMs using epistemic abstention as the primary testbed. Rewriting produces stronger, more persistent, and bidirectional behavioral shifts, including on safety refusal, while reweighting the same examples yields weak, inconsistent effects. The results motivate intervention-aware evaluation of TDA methods.

Hugging Face daily papers·Sep 1, 20:00 UTC · 15d agoAI research

30

Everything in Moderation: Per-Domain Coverage Optima and Alignment-Resistant Domain Gaps in Multi-Domain Mid-Training

Controlled mid-training experiments on Qwen3-8B-Base find each domain has a 10-40% coverage optimum and domain gaps survive alignment SFT.

Using Qwen3-8B-Base (with a 4B replication) across five semantically rule-disjoint KOR-Bench domains, the authors train 30 data allocations spanning the five-domain simplex at five seeds each. All five domains show interior optima in the moderate 10-40% coverage band, and domain gaps persist after a fixed-budget compensatory SFT pass, which raises 116/120 cells yet bridges 0/240 pairs at a 5% threshold. Zero coverage collapses mid-training-only accuracy, though a FineWeb-Edu-only control shows the collapse is partly generic drift. The results argue mid-training data composition requires principled design rather than reliance on later alignment.

arXiv cs.AI / cs.LG / cs.CL·Sep 8, 17:31 UTC · 8d agoAI research1

30

General Quantification of Covariate and Concept Shifts

Paper proposes γ*-concept shifts via entropic optimal transport, deriving estimable generalization bounds unifying covariate and concept shift under distribution shift.

The authors show existing definitions of concept shift break when source and target supports mismatch and propose γ*-concept shifts grounded in entropic optimal transport. They derive a general error bound covering broad loss functions, label spaces and stochastic labeling, plus estimators with concentration guarantees. The resulting DataShifts algorithm quantifies distribution shifts and estimates the error bound in most applications, addressing learning bounds that were previously non-estimable from samples.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:57 UTC · 6d agoAI research1

28

Necessary or Sufficient? Evaluating LLM Explanations With Behavioural Evidence

New framework tests whether LLM-cited explanation factors are necessary or sufficient, finding weak correlation across Claude, GPT, and Gemini models.

An arXiv paper introduces black-box intervention tests measuring whether factors LLMs cite in their explanations are necessary or sufficient for their outputs in agent oversight workflows. Across eight models from the Claude, GPT, and Gemini families, Spearman correlations between cited rankings and measured influence ranged from 0.349-0.354 (advisor recommendation) to 0.431-0.580 (prompt monitoring). Uncited factors scored above the lowest cited factor in up to 57.6% of advisor responses, showing cited top-three factors do not reliably identify the most influential inputs.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 17:37 UTC · 12d agoAI safety & security1

44

Generative Marketing Mix Modeling: A Causal Inference Framework Linking GEO and GEM to Business Impact

Paper proposes Generative Marketing Mix Modeling to causally estimate Generative Engine Optimization and Marketing effects on business outcomes.

The authors develop GMMM, a causal inference framework for measuring how often users see and notice a firm's name in generated answers, which standard marketing data ignore. For GEO it combines repeated generated answers with question counts, shares of generative-system usage and notice probabilities; for GEM it uses sponsored placement records with notice probabilities. The framework compares expected business responses under alternative treatment sequences, establishes identification conditions, and is evaluated on simulated product-recommendation answers in English and Japanese.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:57 UTC · 6d agoAI research

20

VoT: Vision-of-Thought for Unified Multimodal Representation Alignment

Researchers propose Vision-of-Thought (VoT), a discrete visual-planning token layer between VLMs and diffusion transformers improving text-to-image semantic alignment.

VoT introduces a discrete visual-thinking layer between vision-language models and diffusion transformers, letting the VLM act as a multimodal planner that emits tokens describing objects and layouts before pixel generation. A specialized VoT tokenizer is trained with VLM alignment, feature reconstruction, and vector-quantization losses. Experiments show improved semantic alignment and a structured, interpretable interface for controllable generation.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:53 UTC · 9d agoAI research

24

Can Foundation Models Moderate Online Content? Evaluating Instruction- vs. Example-Driven Policy Operationalization

ModerationBench shows foundation models can nearly triple Bluesky's moderation F1 (0.60 vs 0.22), with instruction- and example-driven guidance performing comparably.

Researchers built ModerationBench, a new benchmark of 4,000 manually annotated in-the-wild posts from Bluesky, to test whether foundation models can reliably operationalize content moderation policies. They systematically compare instruction-driven guidance (reasoning from policy precepts) with example-driven guidance (generalizing from precedents) for Vision-Language Models. Both paradigms achieve comparable peak effectiveness, and foundation models nearly triple the F1 of Bluesky's deployed moderation system on Random Posts (0.60 vs 0.22).

arXiv cs.AI / cs.LG / cs.CL·Sep 9, 16:30 UTC · 7d agoAI research1

28