Noisy expert imitation learning requires exponential samples for offline methods but polynomial for a variant of on-policy distillation under a noise condition.
super hub Mixed citations
Phi-4 Technical Report
Mixed citation behavior. Most common role is background (66%).
182
Pith papers citing it
24
external citations · Pith
Background
66% of classified citations
abstract
We present phi-4, a 14-billion parameter language model developed with a training recipe that is centrally focused on data quality. Unlike most language models, where pre-training is based primarily on organic data sources such as web content or code, phi-4 strategically incorporates synthetic data throughout the training process. While previous models in the Phi family largely distill the capabilities of a teacher model (specifically GPT-4), phi-4 substantially surpasses its teacher model on STEM-focused QA capabilities, giving evidence that our data-generation and post-training techniques go beyond distillation. Despite minimal changes to the phi-3 architecture, phi-4 achieves strong performance relative to its size -- especially on reasoning-focused benchmarks -- due to improved data, training curriculum, and innovations in the post-training scheme.
hub tools
citation-role summary
background 19
baseline 9
dataset 1
citation-polarity summary
claims ledger
- abstract We present phi-4, a 14-billion parameter language model developed with a training recipe that is centrally focused on data quality. Unlike most language models, where pre-training is based primarily on organic data sources such as web content or code, phi-4 strategically incorporates synthetic data throughout the training process. While previous models in the Phi family largely distill the capabilities of a teacher model (specifically GPT-4), phi-4 substantially surpasses its teacher model on STEM-focused QA capabilities, giving evidence that our data-generation and post-training techniques go
authors
co-cited works
representative citing papers
EHRNote-ChatQA is the first benchmark for evidence-grounded multi-turn clinical QA over longitudinal discharge summaries, containing 16,072 medical-expert-verified pairs across eight categories and revealing LLM weaknesses in evidence grounding and multi-turn consistency.
Six state-of-the-art LLMs systematically prefer Standard American English over AAE continuations, and a training-free activation steering method reduces this bias 5-20x more than prompting while preserving fluency.
MultiHashFormer enables hash-based autoregression in LMs by encoding tokens as multi-hash signatures, outperforming standard Transformers at 100M-3B scales while keeping parameter count constant for multilingual expansion.
Cliff tokens are single tokens triggering LLM math reasoning failures, identified via adaptive z-test threshold on token potential; a taxonomy and Cliff-DPO optimization yield up to +6.6 accuracy gains.
FlexServe introduces recallable secure memory and NPU to enable cooperative secure LLM inference on mobile devices, reporting 10.05X TTFT speedup over a basic TrustZone strawman.
OmniOPD replaces token-level logit matching in on-policy distillation with Monte Carlo chunk-level semantic verification and a peak-entropy scheduler.
Moral Trolley Arena shows frontier LLMs produce composite moral preferences that are compressed rather than additive functions of calibrated component act strengths across Moral Foundations Theory.
CardioLens is a leakage-resistant CMR testbed of 473k slices and 13k QA pairs showing current MLLMs exhibit a large clinical reality gap with category-collapse failures on real workflows.
ARBITER models reasoning trajectory basins in test-time sampling and uses model-internal signals to correct majority-vote failures, recovering part of the oracle gap on math benchmarks.
A proposer-solver agent pair achieves supervised-level video temporal grounding and fine-grained captioning from 2.5K unlabeled videos via self-reinforcing evolution.
Inducing artificial uncertainty on trivial tasks allows training probes that achieve higher calibration on hard data than standard approaches while retaining performance on easy data.
Develops a causal framework unifying generative AI fairness with standard ML, with new decompositions, identification conditions, and estimators demonstrated on LLM race and gender bias.
LoopUS converts pretrained LLMs into looped latent refinement models via block decomposition, selective gating, random deep supervision, and confidence-based early exiting to improve reasoning performance.
CEAEval is a context-aware evaluation system for speech expressive appropriateness, supported by a new Mandarin dataset with multi-dimensional human annotations and a model that outperforms prior systems.
Standard top-k routers in MoE language models often select suboptimal routes for difficult tokens, and updating only the final router layer raises pass@K on AIME and HMMT benchmarks across multiple models.
ScoringModel raises mean Recall@10 to 52.2 on the FakeWiki provenance benchmark from 35.0 for the best baseline, winning 41 of 45 model-by-condition comparisons and gaining 15.7 points on jailbreak-style queries.
Misrouter enables input-only attacks on MoE LLMs by optimizing queries on open-source surrogates to route toward weakly aligned experts and transferring them to public APIs.
Coral cuts multi-LLM serving costs by up to 2.79x and raises goodput by up to 2.39x on heterogeneous GPUs through adaptive joint optimization and a lossless two-stage decomposition that solves quickly.
TRIBE v2 is a multimodal AI model that predicts human brain activity more accurately than linear encoding models and recovers established neuroscientific findings through in-silico testing.
POSTCONDBENCH is a new multilingual benchmark that evaluates LLM postcondition generation on real code using defect discrimination to assess completeness beyond surface matching.
VoxAfford fuses multi-scale voxel features into MLLM output tokens using cross-attention with a learned compatibility gate to achieve SOTA open-vocabulary 3D affordance detection with ~8% mIoU gain and zero-shot robot transfer.
SENECA uses a novel self-consistent missing mass calculation to improve discrete entropy estimates in small-sample regimes and outperforms alternatives in numerical tests.
Auto-FlexSwitch achieves efficient dynamic model merging by decomposing task vectors into sparse masks, signs, and scalars, then making the compression learnable via gating and adaptive bit selection with KNN-based retrieval.
citing papers explorer
-
Behavior Cloning is Not All You Need: The Optimality of On-Policy Distillation for Noisy Expert Feedback
Noisy expert imitation learning requires exponential samples for offline methods but polynomial for a variant of on-policy distillation under a noise condition.
-
EHRNote-ChatQA: A Benchmark for Evidence-Grounded Multi-Turn Clinical Question Answering over Longitudinal Discharge Summaries
EHRNote-ChatQA is the first benchmark for evidence-grounded multi-turn clinical QA over longitudinal discharge summaries, containing 16,072 medical-expert-verified pairs across eight categories and revealing LLM weaknesses in evidence grounding and multi-turn consistency.
-
LLMs Silently Correct African American English: Auditing and Mitigating Dialect Bias via Activation Steering
Six state-of-the-art LLMs systematically prefer Standard American English over AAE continuations, and a training-free activation steering method reduces this bias 5-20x more than prompting while preserving fluency.
-
MultiHashFormer: Hash-based Generative Language Models
MultiHashFormer enables hash-based autoregression in LMs by encoding tokens as multi-hash signatures, outperforming standard Transformers at 100M-3B scales while keeping parameter count constant for multilingual expansion.
-
Cliff Tokens: Identifying Single-Token Failure Triggers in LLM Mathematical Reasoning
Cliff tokens are single tokens triggering LLM math reasoning failures, identified via adaptive z-test threshold on token potential; a taxonomy and Cliff-DPO optimization yield up to +6.6 accuracy gains.
-
FlexServe: A Fast and Secure LLM Serving System for Mobile Devices with Flexible Resource Isolation
FlexServe introduces recallable secure memory and NPU to enable cooperative secure LLM inference on mobile devices, reporting 10.05X TTFT speedup over a basic TrustZone strawman.
-
OmniOPD: Logit-Free On-Policy Distillation via Speculative Verification
OmniOPD replaces token-level logit matching in on-policy distillation with Monte Carlo chunk-level semantic verification and a peak-entropy scheduler.
-
Every Act Has Its Price: Compressed Moral Composition in Frontier LLMs
Moral Trolley Arena shows frontier LLMs produce composite moral preferences that are compressed rather than additive functions of calibrated component act strengths across Moral Foundations Theory.
-
CardioLens: Revealing the Clinical Reality Gap of MLLMs via Multi-Sequence Cardiac MRI Evaluations
CardioLens is a leakage-resistant CMR testbed of 473k slices and 13k QA pairs showing current MLLMs exhibit a large clinical reality gap with category-collapse failures on real workflows.
-
ARBITER: Reasoning Trajectory Basins and Majority Vote Failures in Test-Time Sampling
ARBITER models reasoning trajectory basins in test-time sampling and uses model-internal signals to correct majority-vote failures, recovering part of the oracle gap on math benchmarks.
-
EvoGround: Self-Evolving Video Agents for Video Temporal Grounding
A proposer-solver agent pair achieves supervised-level video temporal grounding and fine-grained captioning from 2.5K unlabeled videos via self-reinforcing evolution.
-
Inducing Artificial Uncertainty in Language Models
Inducing artificial uncertainty on trivial tasks allows training probes that achieve higher calibration on hard data than standard approaches while retaining performance on easy data.
-
Causal Bias Detection in Generative Artificial Intelligence
Develops a causal framework unifying generative AI fairness with standard ML, with new decompositions, identification conditions, and estimators demonstrated on LLM race and gender bias.
-
LoopUS: Recasting Pretrained LLMs into Looped Latent Refinement Models
LoopUS converts pretrained LLMs into looped latent refinement models via block decomposition, selective gating, random deep supervision, and confidence-based early exiting to improve reasoning performance.
-
Evaluating the Expressive Appropriateness of Speech in Rich Contexts
CEAEval is a context-aware evaluation system for speech expressive appropriateness, supported by a new Mandarin dataset with multi-dimensional human annotations and a model that outperforms prior systems.
-
When Are Experts Misrouted? Counterfactual Routing Analysis in Mixture-of-Experts Language Models
Standard top-k routers in MoE language models often select suboptimal routes for difficult tokens, and updating only the final router layer raises pass@K on AIME and HMMT benchmarks across multiple models.
-
DataDignity: Training Data Attribution for Large Language Models
ScoringModel raises mean Recall@10 to 52.2 on the FakeWiki provenance benchmark from 35.0 for the best baseline, winning 41 of 45 model-by-condition comparisons and gaining 15.7 points on jailbreak-style queries.
-
Misrouter: Exploiting Routing Mechanisms for Input-Only Attacks on Mixture-of-Experts LLMs
Misrouter enables input-only attacks on MoE LLMs by optimizing queries on open-source surrogates to route toward weakly aligned experts and transferring them to public APIs.
-
Coral: Cost-Efficient Multi-LLM Serving over Heterogeneous Cloud GPUs
Coral cuts multi-LLM serving costs by up to 2.79x and raises goodput by up to 2.39x on heterogeneous GPUs through adaptive joint optimization and a lossless two-stage decomposition that solves quickly.
-
A foundation model of vision, audition, and language for in-silico neuroscience
TRIBE v2 is a multimodal AI model that predicts human brain activity more accurately than linear encoding models and recovers established neuroscientific findings through in-silico testing.
-
POSTCONDBENCH: Benchmarking Correctness and Completeness in Formal Postcondition Inference
POSTCONDBENCH is a new multilingual benchmark that evaluates LLM postcondition generation on real code using defect discrimination to assess completeness beyond surface matching.
-
VoxAfford: Multi-Scale Voxel-Token Fusion for Open-Vocabulary 3D Affordance Detection
VoxAfford fuses multi-scale voxel features into MLLM output tokens using cross-attention with a learned compatibility gate to achieve SOTA open-vocabulary 3D affordance detection with ~8% mIoU gain and zero-shot robot transfer.
-
SENECA: Small-Sample Discrete Entropy Estimation via Self-Consistent Missing Mass
SENECA uses a novel self-consistent missing mass calculation to improve discrete entropy estimates in small-sample regimes and outperforms alternatives in numerical tests.
-
Auto-FlexSwitch: Efficient Dynamic Model Merging via Learnable Task Vector Compression
Auto-FlexSwitch achieves efficient dynamic model merging by decomposing task vectors into sparse masks, signs, and scalars, then making the compression learnable via gating and adaptive bit selection with KNN-based retrieval.
-
Evaluating Temporal Consistency in Multi-Turn Language Models
Language models frequently violate temporal scope stability in multi-turn dialogues by drifting toward present-day assumptions even when they possess the correct facts.
-
MORPHOGEN: A Multilingual Benchmark for Evaluating Gender-Aware Morphological Generation
MORPHOGEN is a new multilingual benchmark for testing LLMs on gender-aware morphological generation via rewriting first-person sentences to the opposite gender in French, Arabic, and Hindi.
-
Legal2LogicICL: Improving Generalization in Transforming Legal Cases to Logical Formulas via Diverse Few-Shot Learning
Legal2LogicICL improves accuracy and generalization when mapping legal cases to logical formulas by retrieving balanced diverse exemplars at semantic and structural levels, backed by the new Legal2Proleg dataset.
-
RemoteAgent: Bridging Vague Human Intents and Earth Observation with RL-based Agentic MLLMs
RemoteAgent uses RL fine-tuning on VagueEO to align MLLMs for vague EO intent recognition, handling simple tasks internally and routing dense predictions to tools via Model Context Protocol.
-
BAS: A Decision-Theoretic Approach to Evaluating Large Language Model Confidence
BAS aggregates utility from an answer-or-abstain model across risk thresholds and is uniquely maximized by truthful confidence estimates.
-
The limits of bio-molecular modeling with large language models : a cross-scale evaluation
LLMs perform adequately on bio-molecular classification tasks but remain weak on regression, with hybrid architectures outperforming others on long sequences and fine-tuning hurting generalization.
-
ENEC: A Lossless AI Model Compression Method Enabling Fast Inference on Ascend NPUs
ENEC delivers 3.43X higher throughput than DietGPU and 1.12X better compression ratio than nvCOMP for lossless model weight compression on Ascend NPUs, yielding up to 6.3X end-to-end inference speedup.
-
Graph Property Inference in Small Language Models: Effects of Representation and Reasoning Strategy
Small instruction-tuned language models cannot reliably estimate graph-theoretic properties from textual encodings, though adjacency-list formats and multi-branch reasoning reduce errors relative to edge lists and single-path inference.
-
Integrating Domain-Specialized Language Models with AI Measurement Tools for Deterministic Atomic-Resolution Experimentation
Domain-specialized small language models enable deterministic atomic-resolution scanning probe microscopy control with 99.3% command accuracy, lower computational cost, and better domain performance than larger general models.
-
CLIPoint3D: Language-Grounded Few-Shot Unsupervised 3D Point Cloud Domain Adaptation
CLIPoint3D is the first CLIP-based framework for few-shot unsupervised 3D point cloud domain adaptation that reports 3-16% accuracy gains on PointDA-10 and GraspNetPC-10.
-
Schoenfeld's Anatomy of Mathematical Reasoning by Language Models
ThinkARM abstracts LLM reasoning traces into Schoenfeld episodes and shows that exploration steps correlate with correctness while efficiency methods selectively suppress evaluative feedback.
-
TRANSPORTER: Transferring Visual Semantics from VLM Manifolds
TRANSPORTER generates videos from VLM logits using optimal transport to interpret model predictions on object attributes, actions, and scenes.
-
LayerNorm Induces Recency Bias in Transformer Decoders
Stacked causal self-attention combined with LayerNorm induces recency bias in Transformer decoders, reversing the earlier-token bias seen in attention alone.
-
InterChart: Benchmarking Visual Reasoning Across Decomposed and Distributed Chart Information
InterChart is a new benchmark that reveals steep drops in VLM accuracy when moving from single-chart facts to integrative reasoning over 2-3 related charts, with better performance after decomposing complex charts.
-
ReCogDrive: A Reinforced Cognitive Framework for End-to-End Autonomous Driving
ReCogDrive unifies VLM scene understanding with a diffusion planner reinforced by DiffGRPO to reach state-of-the-art results on NAVSIM and Bench2Drive benchmarks.
-
PRIMETIME : Limits of LLMs in Temporal Primitives
PRIMETIME generator reveals that LLM datetime parsing and arithmetic primitives are individually unreliable but fully learnable via fine-tuning, enabling frontier-level accuracy on event planning with small LoRA models.
-
FlexServe: A Fast and Secure LLM Serving System for Mobile Devices with Flexible Resource Isolation
Page-granular Flex-Mem and switchable Flex-NPU cut TrustZone LLM TTFT by ~10× vs a CMA strawman and ~2.4× vs a pipelined secure-NPU strawman on RK3588.
-
Compress the Cache, Not the Speech Embedding: KV Compression for Efficient Speech LLMs
Learned pooling of speech KV caches from an intermediate LLM layer compresses speech to text-level length while matching or exceeding the uncompressed baseline on ASR and entity recognition, with 1.49–2× decoding speedup.
-
KernelSight-LM: A Kernel-Level LLM Inference Simulator
KernelSight-LM simulates LLM inference at kernel granularity with cross-generation (12.1% per-kernel error) and target-measured (3.8% error) tiers, yielding end-to-end median errors of 15.4%/12.8%/3.0% and 14.3%/6.2%/2.7% for TTFT/TPOT/throughput across six model families.
-
HarmVideoBench: Benchmarking Harmful Video Understanding in Large Multimodal Models
HarmVideoBench is a multi-layered benchmark for harmful video understanding in LVLMs with three hierarchical dimensions, and BCR is a method that raises average model performance from 61.7% to 84.4%.
-
In-Context Model Predictive Generation: Open-Vocabulary Motion Synthesis from Language Models to Physics
ICMPG combines LLM-based candidate generation with MPC-style physical simulation and semantic scoring to produce text-driven human motions that are both plausible and faithful.
-
You Don't Need to Run Every Eval
The benchmark score matrix of 84 models on 133 tasks is approximately rank-2; BenchPress recovers held-out scores to within 4.6 points and identifies 5-benchmark subsets that predict the full scorecard to within 3.93-4.55 points.
-
MINCE: Shrinking LLM Evaluation Datasets via Few-Model Monte Carlo Calibration
MINCE shrinks IFEVAL by 54%, MMLU by 89%, and GSM8K by 70% via few-model Monte Carlo calibration while keeping maximum drift at or below 2.62 percentage points.
-
From Trainee to Trainer: LLM-Designed Training Environment for RL with Multi-Agent Reasoning
The LLM-as-Environment-Engineer framework lets the policy model redesign its own RL environments on the new MAPF-FrozenLake testbed, outperforming larger models and fixed baselines with Qwen3-4B.
-
Multiagent Protocols with Aggregated Confidence Signals
Introduces protocols to aggregate transformed confidence signals from multiagent debates via soft voting or Bayesian fusion, yielding higher AUARC than single agents or standard baselines while keeping F1 stable across benchmarks.
-
Adaptive Multi-Resolution Procedural Knowledge Compression for Large Language Models
SKIM is an adaptive multi-resolution soft-token framework that compresses procedural skills while aiming to preserve logical dependencies and task performance better than prior compression methods.