ProxBench

ProxBench is introduced in the manuscript The Sirens' Song: When Proximal Background Context Overshadows Distant Evidence. The paper identifies the Proximity Trap: distant evidence can be underused not only because it is far from the query, but also because abundant, task-irrelevant background collectively competes for attention.

The benchmark is the paper's controlled evaluation suite for this phenomenon. It is used to compare standard long-context language models with LYRA (Long-context heavY-tailed Relevance Alignment), the attention mechanism proposed in the paper. In this relationship:

• the paper defines and analyzes the Proximity Trap;
• ProxBench measures model robustness under increasingly confusable background; and
• LYRA is evaluated on ProxBench as a proposed mitigation.

The dataset can also be used independently to evaluate other dense-attention, sparse-attention, retrieval-augmented, or context-compression methods. The paper implementation is available in the LYRA github.

Benchmark results

The associated paper evaluates Qwen3-8B, Llama3.1-8B, GLM-4-9B, and LYRA on all four ProxBench levels. Accuracy (%) is reported below; Ours in the original figure denotes LYRA.

ProxBench accuracy across four controlled perturbation levels. Click the figure to open the original PDF.

Model	L1	L2	L3	L4	Average
Qwen3-8B	94.5	82.0	75.0	62.0	78.4
Llama3.1-8B	99.5	61.0	23.0	23.0	51.6
GLM-4-9B	58.5	16.0	2.0	12.0	22.1
LYRA (Ours)	96.0	88.0	80.0	78.0	85.5

LYRA achieves the highest average accuracy at 85.5%, outperforming the strongest evaluated baseline, Qwen3-8B, by 7.1 percentage points. Although Llama3.1-8B is best on L1, LYRA leads on L2, L3, and L4. At the most challenging level, LYRA scores 78.0%, exceeding Qwen3-8B by 16.0 points. Its accuracy decreases by 18.0 points from L1 to L4, compared with drops of 32.5 points for Qwen3-8B, 76.5 points for Llama3.1-8B, and 46.5 points for GLM-4-9B, indicating stronger robustness as the background becomes more confusable.

Dataset description

ProxBench is a controlled synthetic benchmark for testing whether a long-context language model can retrieve and use distant evidence when it must compete with semantically confusable background records. Unlike benchmarks that primarily vary context length, ProxBench holds the retrieval task fixed and systematically changes how strongly the background overlaps with the entity, relation, attribute, semantic role, and value format required by the query.

Each example asks for the access code assigned to a synthetic facility. The relevant evidence appears far from the query in a 64K-token input, while four task-irrelevant records provide controlled interference. A model must return the correct code in the format AA-000 and nothing else. The benchmark therefore measures fine-grained binding and evidence utilization rather than open-ended generation.

This Hugging Face release is evaluation-only. It contains 600 English test examples at a target sequence length of 65,536 tokens. Each input contains 65,504 tokens and reserves 32 tokens for generation.

Task formulation

For a target entity e and a generated access code v, the relevant evidence and query follow the templates:

Evidence: The access code assigned to [e] is [v].
Query:    What is the access code assigned to [e]?
Answer:   [v]

Entities are sampled from an inventory of ten synthetic facility names. Each example contains four background records. The gold code never appears in those records, and no background record establishes the complete binding between the target entity, the queried access-code attribute, and a candidate value.

Perturbation levels

Level	Background design	Capability tested
L1 - Style matched	Uses similar declarative and assignment-oriented syntax, but changes the entity, topic, and answer relation.	Rejecting superficial syntactic similarity.
L2 - Crossed binding	Places the target entity, an access-code cue, another entity, and a candidate code in the same sentence, but binds the code to the other entity.	Resolving argument structure instead of copying a nearby code-shaped span.
L3 - Mixed entity/relation	Alternates between the same access-code relation for another entity and a different attribute for the target entity.	Jointly resolving entity and attribute bindings.
L4 - Fine-grained binding annotations	Retains the mixed binding design and records subtype, entity, relation, attribute, role, and value-format overlap.	Diagnosing the precise source of a binding error.

L3 and L4 use the same text-generation distribution but retain different annotation schemes. They should not be treated as two independent, monotonically increasing text-difficulty levels; L4 is primarily a finer-grained diagnostic view.

Dataset structure

Each level is published as a separate Hugging Face configuration with a single test split.

Configuration	Examples	Target sequence length	Background strategy
L1	200	65,536	style_matched
L2	200	65,536	crossed_binding
L3	100	65,536	mixed_entity_relation
L4	100	65,536	mixed_binding_anchor
Total	600	65,536	-

The proximity-condition field and condition suffixes in row identifiers were removed from the published records. L1 and L2 identifiers were deterministically renumbered so that every row retains a unique id. File sizes, row counts, and SHA-256 checksums are recorded in manifest.json.

Loading the dataset

Replace YOUR_NAMESPACE/ProxBench with the final Hugging Face repository ID.

from datasets import load_dataset

l1 = load_dataset("YOUR_NAMESPACE/ProxBench", "L1", split="test")
all_levels = {
    level: load_dataset("YOUR_NAMESPACE/ProxBench", level, split="test")
    for level in ("L1", "L2", "L3", "L4")
}

A minimal evaluation loop should pass example["input"] to the model and compare the generated text with example["answer"] using the normalization described below.

Data fields

Field group	Fields	Description
Identity	id, pair_id, task, level, level_number	Example and benchmark identifiers.
Model I/O	input, outputs, answer, answer_prefix, metric	Serialized prompt and gold response. outputs contains the acceptable answer list.
Length control	target_sequence_length, tokens_to_generate, length, input_tokens	Sequence-length and generation-budget metadata.
Span control	evidence_span, background_spans, query_span	Token spans for the evidence, four background records, and query.
Position control	evidence_distance, background_distance, background_distance_target, evidence_depth_target	Intended and realized placement metadata.
Perturbation design	background_strategy, background_type, component_strategies	Background construction family and diagnostic subtypes.
Overlap diagnostics	entity_overlap, relation_overlap, attribute_overlap, role_overlap, answer_overlap, key_overlap, candidate_value_overlap, joint_cue_overlap, explicit_negation, distractor_value_format	Structured indicators of how distractors overlap with the target fact.
Similarity diagnostics	evidence_background_lexical_jaccard, query_background_lexical_jaccard	Lexical overlap between evidence/query and background.
Reproducibility	seed, sample_seed, construction_attempt, source_release, source_level	Generation provenance and deterministic retry metadata.

The span fields record the token positions produced and verified during dataset construction. Consumers should use these stored spans rather than attempting to recover evidence with string matching alone.

Evaluation

The primary metric is normalized strict exact match. Before comparison, an optional Answer: prefix, capitalization differences, surrounding whitespace, and a trailing period may be normalized. Any additional explanation should be counted as incorrect. Report accuracy separately for each configuration. If an aggregate score is needed, state whether it is a macro-average over the four levels or a micro-average over all 600 examples, because the configurations have different numbers of examples.

Recommended generation behavior:

Return only the requested access code, for example: AB-123

Data construction and quality control

The release applies automatic constraints during generation:

• No answer leakage: the gold access code is prohibited from appearing in any background record.
• Distinct candidate values: distractor codes differ from the gold answer and from other distractor codes in the same example.
• Entity separation: templates that require another entity exclude the target entity during sampling.
• Binding validity: no background record assigns a distractor value as the access code of the target entity.
• Exact background count: every example contains four recoverable perturbation spans.
• Length control: the tokenized input plus the reserved generation budget matches the target sequence length.
• Span verification: evidence, query, and background spans are verified after serialization and re-tokenization; failed alignments are regenerated.
• Reproducibility: example seeds are deterministically derived from the global seed, context length, perturbation level, and example index. Release shards include byte counts and SHA-256 checksums.

Intended use and limitations

ProxBench is intended for controlled research on long-context retrieval, evidence utilization, binding errors, and robustness to confusable context. It is not a training corpus and should not be interpreted as a broad measure of general long-context reasoning.

The benchmark is synthetic, English-only, and centered on one access-code retrieval template with a narrow answer format. Its regular structure and fixed 64K length do not capture the full diversity of natural documents, multilingual inputs, dialogue, multi-hop reasoning, or open-ended generation. Results should therefore be reported alongside broader long-context benchmarks. Because all entities and records are synthetic, the dataset is not intended to represent real facilities, credentials, or people.