A 0.6B LM with length-aware attention adjustments performs competitive in-context retrieval at million-token scale on MS MARCO, NQ, and LIMIT benchmarks.
hub
Re- trieval head mechanistically explains long-context factu- ality.arXiv preprint arXiv:2404.15574
23 Pith papers cite this work, alongside 1 external citations. Polarity classification is still indexing.
23
Pith papers citing it
1
external citations · Pith
abstract
Despite the recent progress in long-context language models, it remains elusive how transformer-based models exhibit the capability to retrieve relevant information from arbitrary locations within the long context. This paper aims to address this question. Our systematic investigation across a wide spectrum of models reveals that a special type of attention heads are largely responsible for retrieving information, which we dub retrieval heads. We identify intriguing properties of retrieval heads:(1) universal: all the explored models with long-context capability have a set of retrieval heads; (2) sparse: only a small portion (less than 5\%) of the attention heads are retrieval. (3) intrinsic: retrieval heads already exist in models pretrained with short context. When extending the context length by continual pretraining, it is still the same set of heads that perform information retrieval. (4) dynamically activated: take Llama-2 7B for example, 12 retrieval heads always attend to the required information no matter how the context is changed. The rest of the retrieval heads are activated in different contexts. (5) causal: completely pruning retrieval heads leads to failure in retrieving relevant information and results in hallucination, while pruning random non-retrieval heads does not affect the model's retrieval ability. We further show that retrieval heads strongly influence chain-of-thought (CoT) reasoning, where the model needs to frequently refer back the question and previously-generated context. Conversely, tasks where the model directly generates the answer using its intrinsic knowledge are less impacted by masking out retrieval heads. These observations collectively explain which internal part of the model seeks information from the input tokens. We believe our insights will foster future research on reducing hallucination, improving reasoning, and compressing the KV cache.
hub tools
citation-role summary
background 3
citation-polarity summary
roles
background 3polarities
background 3representative citing papers
Replay pairing shows LLM agents do not persist plans in hidden states but rely on plans remaining in context, with rapid signal decay and task performance drops when plans are evicted.
Reroute turns irreversible visual-token pruning into recoverable routing that reuses existing attention scores, improving grounding performance under aggressive reduction on LLaVA-1.5 and Qwen while preserving TFLOPs and KV-cache budgets.
In 1-3B instruction-tuned LMs on GSM8K, arithmetic CoT readout is dominated by positional copying of the trailing number before the answer delimiter, accounting for 54-92 percentage points of accuracy.
SITE applies soft gradient-based head selection to inject ICL-derived task embeddings, outperforming prior embedding adaptation and few-shot ICL across generation, reasoning, and NLU tasks on 12 LLMs from 4B to 70B parameters.
Unsupervised MLP activation dispersion separates known from fabricated entities at AUROC 0.95–1.00 across Bielik scales, while factual reliability scales separately and refusals stay near zero.
Explicitly disentangling semantic and positional streams in a Transformer encoder reveals that absolute positional representations collapse to a 2D document-structure manifold, attention heads specialize by role, and the approach improves linguistic probing performance on 49 of 65 phenomena.
AB-Sparse adaptively allocates per-head block sizes for sparse attention, adds lossless centroid quantization and custom variable-block GPU kernels, and reports up to 5.43% accuracy gain over fixed-block baselines with no throughput loss.
A small set of sparse autoencoder features in LLMs drives shifts between generous and selfish allocations in dictator games, with causal patching and steering confirming their role and generalization to other social games.
Stylistic rewrites of harmful prompts raise attack success rates from 3.84% to 36.8-65% across 31 frontier models, indicating weak generalization in safety refusals.
SAGE is a training-free context reduction method that converts attention signals from a small LLM into a differential relevance heatmap to select top units for downstream QA, achieving competitive accuracy at 10% token budget on benchmarks like QuALITY-hard.
XComp reaches extreme video compression (one token per selective frame) via learnable progressive token compression and question-conditioned frame selection, lifting LVBench accuracy from 42.9 percent to 46.2 percent after tuning on 2.5 percent of standard data.
A single dense-pretrained RAT+ model can be adapted with 1B tokens to run at dilation sizes D=2..128, matching dense accuracy at D=16 and losing only 1-3 points at D=64 on reasoning and long-context benchmarks, though retrieval tasks drop far more.
BLASST dynamically sparsifies attention by thresholding softmax scores to skip blocks, delivering 1.5x speedups at 70%+ sparsity while preserving benchmark accuracy.
PyramidKV dynamically compresses KV cache across layers following pyramidal information funneling, matching full performance at 12% retention and outperforming alternatives at 0.7% retention with up to 20.5 accuracy gains.
RaBitQCache proposes rotated binary quantization with binary-INT4 arithmetic for unbiased attention weight estimation in long-context LLMs, enabling adaptive Top-p retrieval and hardware optimizations.
Reasoning in large output spaces proceeds via shortlisting then fine-grained reasoning; this characterization enables a mechanistic distillation strategy that outperforms standard distillation.
DCO is an inference-time intervention that decomposes attention head outputs orthogonally to a dynamic context anchor and suppresses outlier components via Z-score to improve contextual faithfulness in Llama models.
A training-free method that steers specific attention heads toward caption-query states reduces object hallucination in LVLMs by ~6% on average across five benchmarks.
Flux Attention uses a context-aware Layer Router to dynamically assign full or sparse attention to each LLM layer, achieving up to 2.8x prefill and 2.0x decode speedups with competitive performance on long-context and reasoning tasks.
ART replaces uniform attention in shallow LLM layers with local attention patterns to reduce hallucinations across multiple model architectures.
Enforcing sentence-level citations degrades LLM attribution quality by 16-276% versus paragraph-level, with larger models penalized more due to disrupted semantic synthesis.
A survey-style reference book mapping the full agentic-AI stack from transformer internals to production deployment, with no new research result.
citing papers explorer
-
Can Language Models Actually Retrieve In-Context? Drowning in Documents at Million Token Scale
A 0.6B LM with length-aware attention adjustments performs competitive in-context retrieval at million-token scale on MS MARCO, NQ, and LIMIT benchmarks.
-
Plans Don't Persist: Why Context Management Is Load Bearing for LLM Agents
Replay pairing shows LLM agents do not persist plans in hidden states but rely on plans remaining in context, with rapid signal decay and task performance drops when plans are evicted.
-
Reroute, Don't Remove: Recoverable Visual Token Routing for Vision-Language Models
Reroute turns irreversible visual-token pruning into recoverable routing that reuses existing attention scores, improving grounding performance under aggressive reduction on LLaVA-1.5 and Qwen while preserving TFLOPs and KV-cache budgets.
-
The Readout Shortcut: Positional Number Copying Dominates Arithmetic CoT Readout in Small Language Models
In 1-3B instruction-tuned LMs on GSM8K, arithmetic CoT readout is dominated by positional copying of the trailing number before the answer delimiter, accounting for 54-92 percentage points of accuracy.
-
Soft Head Selection for Injecting ICL-Derived Task Embeddings
SITE applies soft gradient-based head selection to inject ICL-derived task embeddings, outperforming prior embedding adaptation and few-shot ICL across generation, reasoning, and NLU tasks on 12 LLMs from 4B to 70B parameters.
-
Does Bielik Know What It Doesn't Know? Activation Dispersion Separates Entity Familiarity from Factual Reliability Across Model Scale
Unsupervised MLP activation dispersion separates known from fabricated entities at AUROC 0.95–1.00 across Bielik scales, while factual reliability scales separately and refusals stay near zero.
-
Give it Space! Explicit Disentangling of Positional and Semantic Representations in Encoders
Explicitly disentangling semantic and positional streams in a Transformer encoder reveals that absolute positional representations collapse to a 2D document-structure manifold, attention heads specialize by role, and the approach improves linguistic probing performance on 49 of 65 phenomena.
-
AB-Sparse: Sparse Attention with Adaptive Block Size for Accurate and Efficient Long-Context Inference
AB-Sparse adaptively allocates per-head block sizes for sparse attention, adds lossless centroid quantization and custom variable-block GPU kernels, and reports up to 5.43% accuracy gain over fixed-block baselines with no throughput loss.
-
Understanding the Mechanism of Altruism in Large Language Models
A small set of sparse autoencoder features in LLMs drives shifts between generous and selfish allocations in dictator games, with causal patching and steering confirming their role and generalization to other social games.
-
Adversarial Humanities Benchmark: Results on Stylistic Robustness in Frontier Model Safety
Stylistic rewrites of harmful prompts raise attack success rates from 3.84% to 36.8-65% across 31 frontier models, indicating weak generalization in safety refusals.
-
SAGE: Selective Attention-Guided Extraction for Token-Efficient Document Indexing
SAGE is a training-free context reduction method that converts attention signals from a small LLM into a differential relevance heatmap to select top units for downstream QA, achieving competitive accuracy at 10% token budget on benchmarks like QuALITY-hard.
-
One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding
XComp reaches extreme video compression (one token per selective frame) via learnable progressive token compression and question-conditioned frame selection, lifting LVBench accuracy from 42.9 percent to 46.2 percent after tuning on 2.5 percent of standard data.
-
RAT+: Train Dense, Infer Sparse -- Recurrence Augmented Attention for Dilated Inference
A single dense-pretrained RAT+ model can be adapted with 1B tokens to run at dilation sizes D=2..128, matching dense accuracy at D=16 and losing only 1-3 points at D=64 on reasoning and long-context benchmarks, though retrieval tasks drop far more.
-
BLASST: Dynamic BLocked Attention Sparsity via Softmax Thresholding
BLASST dynamically sparsifies attention by thresholding softmax scores to skip blocks, delivering 1.5x speedups at 70%+ sparsity while preserving benchmark accuracy.
-
PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funneling
PyramidKV dynamically compresses KV cache across layers following pyramidal information funneling, matching full performance at 12% retention and outperforming alternatives at 0.7% retention with up to 20.5 accuracy gains.
-
RaBitQCache: Rotated Binary Quantization for KVCache in Long Context LLM Inference
RaBitQCache proposes rotated binary quantization with binary-INT4 arithmetic for unbiased attention weight estimation in long-context LLMs, enabling adaptive Top-p retrieval and hardware optimizations.
-
Characterize Then Distill: Mechanistic Reasoning in Large Output Spaces
Reasoning in large output spaces proceeds via shortlisting then fine-grained reasoning; this characterization enables a mechanistic distillation strategy that outperforms standard distillation.
-
Hallucinations as Orthogonal Noise: Inference-Time Manifold Alignment via Dynamic Contextual Orthogonalization
DCO is an inference-time intervention that decomposes attention head outputs orthogonally to a dynamic context anchor and suppresses outlier components via Z-score to improve contextual faithfulness in Llama models.
-
CAST: Mitigating Object Hallucination in Large Vision-Language Models via Caption-Guided Visual Attention Steering
A training-free method that steers specific attention heads toward caption-query states reduces object hallucination in LVLMs by ~6% on average across five benchmarks.
-
Flux Attention: Context-Aware Hybrid Attention for Efficient LLMs Inference
Flux Attention uses a context-aware Layer Router to dynamically assign full or sparse attention to each LLM layer, achieving up to 2.8x prefill and 2.0x decode speedups with competitive performance on long-context and reasoning tasks.
-
ART: Attention Replacement Technique to Improve Factuality in LLMs
ART replaces uniform attention in shallow LLM layers with local attention patterns to reduce hallucinations across multiple model architectures.
-
Are Finer Citations Always Better? Rethinking Granularity for Attributed Generation
Enforcing sentence-level citations degrades LLM attribution quality by 16-276% versus paragraph-level, with larger models penalized more due to disrupted semantic synthesis.
-
The Hitchhiker's Guide to Agentic AI: From Foundations to Systems
A survey-style reference book mapping the full agentic-AI stack from transformer internals to production deployment, with no new research result.