SELD Detects Silent Evidence Loss in Context Compression

Summary:Context compression can remove the evidence needed to justify an answer while preserving the model's final output, creating a silent failure mode that answer-level checks cannot detect. SELD addresses this problem without training or labeled examples by comparing how a reader model ranks unspoken candidate answers before and after compression. The accompanying SEL-Pairs benchmark contains length-matched evidence-removing and control deletions from three QA datasets. Across five reader models, m

---

The blind spot of unchanged answers

Long-context pipelines routinely summarize, select, or compress input before handing it to a reader model. Most quality checks compare the final answer, top predicted tokens, or perplexity. These checks miss a structural failure mode: the evidence supporting an answer may be removed while the reader still produces the same output.

This silent evidence loss is particularly dangerous. The current model may still guess correctly, but a compressed context cached in external memory or reused by another model no longer has an auditable basis for that answer. SELD therefore asks whether the reader's preference structure over unspoken candidates changed, rather than whether its top answer changed.

How SELD detects hidden damage

The central observation is that deleting evidence can leave the preferred answer untouched while shifting the relative order of alternative answers. Comparing candidate-answer preferences before and after compression can therefore expose failures that answer-level checks miss.

A practical implementation would:

1. Run the reader model on both the original and compressed contexts.
2. Collect candidate answers and their conditional probabilities or scores, including candidates not explicitly emitted.
3. Align the candidate sets and compare their rankings, distributions, or deviations from the original context.
4. Flag cases where the final answer remains stable but the preference distribution changes materially.

The abstract does not specify how candidates are generated, how scores are normalized, or which statistical tests are used. The reported 83% should therefore not be treated as a universal production accuracy. Its likely cost burden is extra sampling and scoring rather than training: detection requires access to the reader's candidate distribution, and APIs that expose only limited outputs may weaken the signal.

Why length-matched control deletions matter

SEL-Pairs constructs evidence removals and control deletions from three QA datasets while matching their deletion lengths. This reduces a major confound: a detector might react to shorter input, positional shifts, or generic fluency changes rather than the removal of answer-bearing evidence.

Length-matched controls create a stricter counterfactual test. If evidence removals induce a different signal from ordinary content removals, the method is more likely to detect evidence deletion rather than generic distribution shift. The dataset composition, compression algorithms, compression-ratio curves, and full ablation results remain necessary for judging generalization.

Results and evaluation limits

The experiments cover five reader models, multiple compressors, and several compression ratios. According to the abstract, SELD identifies the evidence-removing edit in 83% of matched pairs whose answers remain unchanged, compared with 70% for the strongest baseline, an absolute improvement of 13 percentage points.

The result indicates that candidate-preference comparison captures information invisible to answer checks, but several evaluation questions remain:

• Paired detection versus open-world auditing: Identifying a known edit in a matched pair does not mean the system can determine whether an arbitrary compressed context is safe to reuse.
• Baseline strength: The comparison should include semantic-entailment, evidence-attribution, perturbation-sensitivity, and repeated-sampling methods, not merely answer agreement.
• Asymmetric error costs: A false negative passes unsupported context downstream; a false positive may block valid compression or trigger an expensive retry. Precision, recall, and the cost per million tokens should be reported together.
• Model dependence: Candidate ranking and calibration quality depend on the reader model. Stability across model families, prompt templates, and tasks needs independent reproduction.

Deployment recommendations

SELD is best positioned as an evidence-integrity gate in a context pipeline, not as a replacement for the compressor itself. Teams can sample compressed outputs for high-risk knowledge bases, shared agent memory, and long-document question answering.

A practical workflow is:

1. Preserve original evidence snippets or content-addressed references so compressed context remains traceable.
2. Run the reader on the compressed output and export the top-k candidates and probabilities.
3. Measure preference-distribution drift between the original and compressed inputs.
4. For high-risk cases, fall back to the source, widen retrieval, or generate a cited evidence summary.
5. Log false positives, evidence fallbacks, end-to-end task accuracy, and Token cost in production.

For cached compression, detection results should be bound to the compression version, reader model, and prompt template. Changing the model, quantization scheme, or prompt can invalidate earlier conclusions.

Cost and deployment trade-offs

SELD requires neither training nor human labels, which is its main advantage over supervised evidence verifiers. The trade-off is runtime overhead: candidate distributions must be collected, and separate forward passes for original and compressed contexts may be required. If candidate generation shares computation with the reader call, incremental cost may be modest; if multiple samples are needed, cost grows with context length and candidate count.

Before adoption, teams should measure overhead per million Tokens, the interception rate for silent evidence loss, and the cost of retries caused by false positives. The method has clear economic value when one unsupported context can contaminate long-term memory, propagate to downstream models, or trigger a costly downstream decision.

---

Source: ArXiv NLP/LLM (cs.CL) (2026-11-11)

— END —

Tags: #context compression Follow #silent evidence loss Follow #long context Follow #model evaluation Follow #evidence integrity Follow