Distracted Alignment: Stress-Testing Protein–Text Models with Generated Noise

Models that read a protein's amino-acid chain and write an English description of its function are built by bolting a protein encoder onto a language model through a small "projector" — a learned adapter that translates protein representations into tokens the language model can attend to. Every such system to date — ProtT3, Prot2Text-V2, BioBridge — is trained and evaluated on clean, perfectly paired data: one protein, its correct curated description, nothing else. Real biological prompts are never like this. They contain wrong annotations, irrelevant literature about distantly related proteins, and sequences with experimental artifacts. Two separate literatures predict trouble but have never been brought together. Text-only models collapse when distractors are injected into their prompts — an automated tree search for model-specific distractions caused roughly 45% average accuracy drops across frontier LLMs (Breaking Focus). And general LLMs fed raw biological sequence tokens alongside text actually reason worse than with text alone: on a protein question-answering benchmark, Gemini 2.5 Pro scored 87.19 with textual context only, 18.81 with the sequence only, and degraded below 87.19 when both were given (Lost in Tokenization).

This leaves an unmeasured question at the center of the alignment architecture: does the projector itself act as a noise channel? Nobody knows whether projected protein representations distract the language model's attention the way raw sequence tokens do, whether domain-specific fine-tuning makes these small specialized models more fragile rather than less, and whether the clean static benchmarks used to grade these systems could even detect any of it, since sequence-identity clustering does not stop functional text from leaking into pretraining corpora (PFMBench). This proposal builds the missing measurement instrument, runs the diagnosis, and then tests where robustness has to be trained into the system.

Core Claim

The claim is threefold. First, current protein–text models are systematically fragile under contextual distraction, but this fragility is invisible because every existing evaluation is clean, static, and homology-split rather than distraction-tested and contamination-audited. Second, the fragility is modality-asymmetric in a way we can quantify: noise in the textual prompt and noise in the sequence/projector channel should hurt differently, mirroring the text-versus-image asymmetry documented for vision-language models, where misleading text hints dropped InternVL2-2B by 9.05% while irrelevant images caused near-zero damage (On the robustness of multimodal language model towards distractions). Third, robustness cannot simply be patched onto the decoder: it must be (or need not be — this is the decisive comparison) trained into the alignment stage itself, by exposing the projector to corrupted pairings during training rather than only clean ones.

The proposal deliberately treats the four research directions asymmetrically. Cross-modal alignment is the object under study; contextual noise sensitivity supplies the failure model and the attack machinery; benchmark contamination is a methodological constraint that dictates how the evaluation must be built; and domain-specific fine-tuning is the regime where we expect the pathology to be amplified, via the "capability curse" — the finding that more capable small language models internalize corrupted fine-tuning data more readily than weaker ones, with near-total task collapse already at 25% syntactic contamination (Sensitivity of Small Language Models to Fine-tuning Data Contamination).

Research Program

The first deliverable is BioDistract, a dynamically generated evaluation suite for protein–text models. It adapts two existing generation ideas to the biological modality. From Breaking Focus we take error-guided perturbation: a tree search that proposes distractors, keeps the ones that actually fool a held-out model, and discards weak ones — this is what separates genuinely damaging noise from noise a model shrugs off, and it is why their tree search induced a 52% drop where a generic perturbation baseline (DyVal2) induced only 15.2%. From DyVal we take the principle that evaluation instances should be generated with controllable difficulty knobs rather than drawn from a fixed static set, since static-benchmark leaders like phi-1.5 and WizardMath-13B scored nearly zero on dynamically generated equivalents. In BioDistract, the base instance is a protein (from a time-split slice of Swiss-Prot) plus a question about its function, localization, or pathway, and the controllable distractors come in two modality channels: textual ones (a plausible but wrong Gene Ontology annotation, the description of a non-homologous protein retrieved by keyword overlap, an irrelevant literature snippet) and sequence ones (point mutations in functional residues, shuffled domains, insertion of an unrelated segment passed through the protein encoder). Difficulty knobs are distractor count, homology distance of the distracting protein, and how many residues are perturbed.

Contamination control is built in from the start, not bolted on afterward. Instances are drawn preferentially from proteins first characterized in recent years — Lost in Tokenization showed sequence-based models collapse on proteins discovered between 2014 and 2024 while text-context approaches degrade gracefully, which is exactly the signature of memorized historical associations rather than generalization. Beyond temporal splitting, we audit each base instance with token-level surprise and pretraining-corpus frequency probes in the spirit of Snoopy, flagging any instance whose answer text appears in likely pretraining data. We report both a filtered and unfiltered score, so the field can see how much of current protein–text performance survives decontamination. This matters because the identity-threshold clustering used by Prot2Text-V2 (40%) and PFMBench (30%) blocks sequence homology leakage but not functional-text leakage.

The second deliverable is the diagnosis, run across three model families that represent the three architectural strategies: projector-based generative models (ProtT3, Prot2Text-V2, BioBridge), raw-sequence instruction-tuned LLMs (ChatMultiOmics, fine-tuned per Mol-Instructions), and general LLMs prompted with text context only (the strongest simple strategy, per Lost in Tokenization). For each we measure a distraction vulnerability score

$$D = \frac{A_{\text{clean}} - A_{\text{distracted}}}{A_{\text{clean}}},$$

where $A$ is task accuracy on generated QA or caption quality (BLEU/ROUGE against the curated reference) on captioning instances, computed separately for text-channel and sequence-channel distractors. We additionally run the capability-curse experiment: fine-tune the same base LLMs (Llama-3.1-8B, Qwen2.5-7B, and smaller family members) on instruction sets containing 25%–100% corrupted functional annotations, following the four contamination types of Sensitivity of Small Language Models to Fine-tuning Data Contamination, and test whether the more capable fine-tuned models become more fragile on BioDistract.

The third deliverable is the mitigation comparison, which is deliberately framed as a question rather than a bet: does robustness need to be trained into the alignment stage, or can it be patched at the decoder? The candidate method trains the projector with noise-augmented alignment — for every clean Swiss-Prot pairing, the training batch also contains a corrupted variant (distracting text, or a perturbed sequence), and the alignment objective must score the correct pairing above the corrupted one, so the projector learns which modality to trust when they disagree. The competitor is the decoder-side fix: standard clean alignment followed by Direct Preference Optimization — a lightweight post-training method that teaches a model to prefer clean-context answers over distracted-context answers — exactly the recipe that restored robustness in Breaking Focus. Both get the same base model, the same data budget, and the same tuning compute; we also report the combination.

Baselines

The naive reference point is doing nothing beyond existing practice: a projector-based protein–text model evaluated on clean prompts. The concrete instantiation is Prot2Text-V2, which pairs a frozen ESM2-3B protein encoder with a Llama-3.1-8B-Instruct decoder through a lightweight projector trained with hybrid contrastive alignment, and which reports a BLEU-4 of 43.34 on its Swiss-Prot test split at under 40% sequence identity, against 32.45 for the original Prot2Text (Prot2Text-V2). This is the number our models must not regress on: a method that buys robustness by dropping clean-prompt caption quality below roughly 43 BLEU-4 is not a win, it is a trade. Beating this baseline means holding clean-prompt performance within a small margin of 43.34 while substantially reducing the distraction vulnerability $D$ on BioDistract.

The strongest simple competitor is the context-only pipeline: a general LLM prompted with textual summaries from classical bioinformatics tools and no sequence input at all. Its power is documented — Gemini 2.5 Pro scored 87.19 on the Protein QA benchmark with text context only (Lost in Tokenization) — and it requires no training whatsoever. It will be run on the identical BioDistract instances with the same textual distractors injected into its context, so the comparison is exactly matched on data. Beating it requires two things jointly: matching or exceeding its robustness under text-channel distraction, and outperforming it in the low-homology regime where homology-based summaries bottom out, the regime where Prot2Text-V2 showed aligned models still produce sound descriptions for proteins with BLAST bitscores of 0–30 and where expert blind reviewers ranked aligned-model outputs above GPT-4o's.

The strongest existing defense method is the DPO-on-distracted-pairs recipe from Breaking Focus, which was shown to restore robustness against distractions that caused about 45% average drops in frontier LLMs. It will be applied to the same base protein–text model, with preference pairs constructed from BioDistract's clean and distracted versions of the same instance, under the same compute and tuning budget as our noise-augmented alignment — neither method gets more data, more steps, or a larger model. Beating it means a smaller $D$ at equal or better clean performance; if DPO matches noise-augmented alignment, that is itself the headline finding, since it would show robustness can be patched at the decoder without touching the projector.

The key ablation isolates the proposal's central ingredient: the identical architecture (same encoder, projector, decoder) trained with standard clean-pair alignment only, versus the same training plus corrupted-pair contrastive terms. The difference between these two runs on BioDistract, with everything else held fixed, is the direct measurement of whether training-time noise exposure is what buys robustness, rather than any incidental effect of data scale or longer training. We will also run a decoder-only variant in which the projector sees only clean data and DPO sees the noise, to locate where the robustness actually lives.

Why not just…?

Why not just prompt the LLM with bioinformatics text summaries and skip cross-modal models entirely? This is the most serious alternative, and the evidence for it is strong: 87.19 for context-only Gemini 2.5 Pro versus 18.81 for sequence-only is not a close race (Lost in Tokenization). But it has a documented failure floor: for orphan or low-homology proteins where similarity search returns nothing informative, Prot2Text-V2 maintained sound caption generation at BLAST bitscores of 0–30 and experts preferred its descriptions over GPT-4o's in blind review. More importantly for this proposal, the context-only pipeline inherits the same distraction pathology — its text context can be corrupted just as easily — and no one has measured that either. So the alternative is included as a first-class baseline rather than argued away: if it wins on BioDistract including the low-homology stratum, the conclusion follows and the cross-modal program is in deeper trouble than its literature admits.

Why not just add "ignore the irrelevant information" to the prompt? Because it has been tested and it fails: Breaking Focus found such instructions completely ineffective against adaptively searched distractions, and On the robustness of multimodal language model towards distractions found defense prompts sometimes made accuracy worse. We will still report the prompted-defense number on BioDistract as a cheap lower bound, but the prior evidence gives no reason to expect it to help, and the proposal's contribution does not depend on it.

Why not just apply DPO on distracted pairs and call it done? This is the closest existing work and a genuine threat to the mitigation component: Breaking Focus already demonstrated the recipe works on text benchmarks, and it is simpler than retraining alignment. The honest answer is that it may work as well, and the proposal is structured so that this outcome is a finding, not a failure: DPO is run head-to-head under matched budgets, and the scientific question it settles — whether the projector is a noise channel that must be corrected where noise enters, or whether decoder preferences suffice — is answered either way. What DPO has never faced is sequence-channel noise, since it was developed for unimodal text; if the projector does inject distractor-like representations, preference pairs that never expose the projector to corrupted sequences may not teach it to gate them.

Why not just use a bigger model and stricter CD-HIT splits? Scale is refuted twice in the file: scaling ESM-2 from 150M to 15B gave only marginal gains, with the 650M multimodal ProTrek winning 75–88% of comparisons and beating ESM2-15B on 6 of 8 core tasks (PFMBench), and frontier models still fell from 85.7% to 22.0% on CommonsenseQA under adaptive distraction (Breaking Focus). On contamination, CD-HIT clustering at 30–40% identity blocks sequence homology leakage but not functional-text leakage (PFMBench), and temporal collapse persists even on clustered benchmarks (Lost in Tokenization). BioDistract's generated instances and temporal slices are designed precisely because neither knob closes the gap.

Proposed Experiments and Decisive Checks

Experiment one measures baseline fragility. All three model families are evaluated on BioDistract with text-channel and sequence-channel distractors separately, on both the temporal slice (pre-2014 versus 2014–2024 discovery dates) and the decontaminated instance set. The decisive readout is the modality asymmetry: if $D_{\text{text}} \gg D_{\text{sequence}}$ across projector models, the projector is functioning as a filter and the pathology is textual; if sequence-channel noise causes large drops in projector models but not in context-only models, the projector is the noise channel, which is the paper's central hypothesis. If neither holds — if protein–text models are simply robust — the hypothesis is refuted and the benchmark plus negative result is still the deliverable, because it would show the fragility literature does not transfer to this modality.

Experiment two runs the capability-curse test in the biological setting: fine-tuned specialized models at several capability levels, with corrupted fine-tuning corpora at graded contamination rates, evaluated on clean and distracted BioDistract. The prediction from Sensitivity of Small Language Models to Fine-tuning Data Contamination is that stronger models internalize the corruptions faster; the open question is whether biological fine-tuning compounds distraction fragility on top of that, and whether general-language preservation techniques like the domain-incremental pretraining of BioBridge (which held MMLU at 63.30 and RACE at 84.00) also protect against distraction or only against forgetting.

Experiment three is the mitigation horse race: noise-augmented alignment versus clean alignment plus DPO versus the combination, all on the same base stack, all evaluated on clean Swiss-Prot captioning (against the 43.34 BLEU-4 reference point), on BioDistract at every difficulty knob, and on the low-homology stratum. Success for the proposal's ingredient means the noise-augmented run beats the DPO-only run on sequence-channel distraction at matched clean performance; success for DPO means robustness is a decoder property and the projector can stay naively trained. Either outcome changes how the next generation of protein–text models should be built, and the contamination audit贯穿 the whole program ensures whatever number ends up in the abstract was earned rather than remembered.